ABSTRACT
Title of dissertation: LAB-ON-CMOS SENSORS AND REAL-TIME
IMAGING FOR BIOLOGICAL CELL MONITORING
Bathiya Prashan Bandara Senevirathna
Doctor of Philosophy, 2019
Dissertation directed by: Professor Pamela Abshire
Department of Electrical & Computer Engineering
Monitoring biological cell growth and viability is essential for in vivo biomedi-
cal diagnosis and therapy, and in vitro studies of pharmaceutical efficacy and mate-
rial toxicity. Conventional monitoring techniques involve the use of dyes and markers
that can potentially introduce side effects into the cell culture and often function as
end-point assays. This eliminates the opportunity to track fast changes and to de-
termine temporal correlation between measurements. Particularly in drug screening
applications, high-temporal resolution cell viability data could inform decisions on
drug application protocols that could lead to better treatment outcomes.
This work presents development of a lab-on-chip (LoC) sensor for real-time
monitoring of biological cell viability and proliferation, to provide a comprehensive
picture of the changes cells undergo during their lifecycle. The LoC sensor con-
sists of a complementary metal-oxide-semiconductor (CMOS) chip that measures
the cell-to-substrate coupling of adherent cells that are cultured directly on top.
This technique is non-invasive, does not require biochemical labeling, and allows for
automated and unsupervised cell monitoring. The CMOS capacitance sensor was
designed to addresses the ubiquitous challenges of sensitivity, noise coupling, and
dynamic range that affect existing sensors. The design includes on-chip digitization,
serial data output, and programmable control logic in order to facilitate packaging
requirements for biological experiments. Only a microcontroller is required for read-
out, making it suitable for applications outside the traditional laboratory setting.
An imaging platform was developed to provide time-lapse images of the sensor sur-
face, which allowed for concurrent visual and capacitance observation of the cells.
Results showed the ability of the LoC sensor to detect single cell binding events
and changes in cell morphology. The sensor was used in in vitro experiments to
monitor chemotherapeutic agent potency on drug-resistant and drug-sensitive can-
cer cell lines. Concentrations higher than 5 ?M elicited cytotoxic effects on both cell
lines, while a dose of 1 ?M allowed discrimination of the two cell types. The system
demonstrates the use of real-time capacitance measurements as a proof-of-concept
tool that has potential to hasten the drug development process.
LAB-ON-CMOS SENSORS AND REAL-TIME IMAGING FOR
BIOLOGICAL CELL MONITORING
by
Bathiya Senevirathna
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2019
Advisory Committee:
Professor Pamela Abshire, Chair/Advisor
Professor Reza Ghodssi
Professor Timothy Horiuchi
Professor Robert Newcomb
Professor Ian White
Professor Elisabeth Smela*
?c Copyright by
Bathiya Senevirathna
2019
Dedication
To my parents.
ii
Acknowledgments
I would first like to express my gratitude to my advisor, Dr. Pamela Abshire.
I thank her for giving me a chance to work on this exciting research, for her patient
guidance throughout all aspects of the research, and for being such a wonderful
mentor. Her infectious enthusiasm and positive attitude makes everyone around her
a better person. A special thanks as well to Dr. Elisabeth Smela for her guidance,
and valuable and constructive suggestions during the planning and development of
this work, as well as the resources for performing experiments.
I would like to thank Dr. Pamela Abshire, Dr. Reza Ghodssi, Dr. Timothy
Horiuchi, Dr. Robert Newcomb, Dr. Elisabeth Smela?, and Dr. Ian White for
graciously agreeing to serve on my dissertation committee and sparing the time to
review the work.
I would of course like to thank all my lab mates at the Integrated Biomorphic
Information Systems (IBIS) Laboratory. Alex Castro for wide-ranging discussions,
collaborations for coursework, and general advice throughout my entire graduate
student life. Sheung Lu for his invaluable help on this and other projects. From
working on chip packaging, to the discussions on eccentric ideas, and the help set-
ting up and performing experiments, this dissertation could not have been realized
without his help. Marc Dandin for the valuable technical discussions, logistical help
in moving the project forward, and general life advice. His influence at IBIS cannot
be overstated. Timir Datta for the early mentorship and allowing me to get involved
in the projects that eventually led me to this work. Tsung-Hsueh Lee and Andrew
?I would like to thank Dr. Elisabeth Smela for also being a part of my committee, although
she was unable to attend in person due to last minute scheduling conflicts.
iii
Berkovich for advice and assistance setting up CAD tools for VLSI design. Emily
Naviasky, Somashekar Prakash, and Nicole Nelson for laying the groundwork for
this project.
I am particularly grateful to our collaborators from the University of Oulu
and Linko?ping University, for the technical discussions in developing this project.
In particular, Joni Kilpija?rvi and Dr. Anita Lloyd Spetz for assembling devices
that were crucial for obtaining some of the experimental data presented in this
dissertation. I would additionally like to thank Dr. John Basile for the valuable
insight on this project from a biologists perspective, and Dr. Yan Shu for graciously
providing the cells used in experiments, both from the University of Maryland,
Baltimore.
I would also like to express my appreciation to Dr. Ricardo Araneda, Dr.
Jonathan Simon, Dr. Bradley Hatfield, Adriane Fang, Karen Bradley for the lively
dialogue on other collaborations. Jooik Chung, Shashikant Koul, Jaime Campos,
and honorary IBIS member, Estefany Carillo, thank you for the broad discussions
and being a positive influence in grad school. Diane Woods and Kirubel Mersha,
thank you for the impetus to go to graduate school and inspiration.
I would also like to thank numerous staff at ECE/ISR/IT, including Jeff McK-
inney, Gwen Flasinski, Jay Renner, and Regina King for support resources, and Dr.
Melanie Prange for the academic support and valuable advice during my time at
UMD.
Finally, I?d like to thank my parents, aunts, uncles, and cousins for their
continuous support and words of encouragement.
iv
Table of Contents
List of Tables viii
List of Figures viii
List of Abbreviations xiii
1 Introduction 1
1.1 Lab-on-CMOS Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Capacitance Sensing on CMOS 6
2.1 Sensing Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Charge based capacitance measurements . . . . . . . . . . . . 6
2.1.1.1 Fingerprint Sensor . . . . . . . . . . . . . . . . . . . 7
2.1.1.2 Metal Interconnect Characterization . . . . . . . . . 8
2.1.2 Frequency-based Capacitance Measurements . . . . . . . . . . 9
2.2 Cell Adhesion & and Viability sensing . . . . . . . . . . . . . . . . . 11
2.2.1 Cell-Substrate Coupling . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Non-Electronic Techniques . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Electric Cell-Substrate Impedance Sensing . . . . . . . . . . . 14
2.2.4 Integrated Sensing on CMOS . . . . . . . . . . . . . . . . . . 15
2.2.5 Discussion of Integrated CMOS Approaches . . . . . . . . . . 16
2.2.6 This Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Capacitance Sensor Design 21
3.1 Cell Capacitance Model . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Multi-physics Simulations . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Capacitive Sensing Approach . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Sensor Element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4.1 Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4.2 Simulation Performance . . . . . . . . . . . . . . . . . . . . . 28
3.4.2.1 Capacitance Sensitivity . . . . . . . . . . . . . . . . 28
3.4.2.2 Temperature Sensitivity . . . . . . . . . . . . . . . . 29
3.5 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5.1 Multiplexer and Counter . . . . . . . . . . . . . . . . . . . . . 31
3.5.2 Serial Readout . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.6 Fabricated Chip & Readout Platform . . . . . . . . . . . . . . . . . . 34
3.7 Capacitance Computation . . . . . . . . . . . . . . . . . . . . . . . . 36
3.8 Bench-Top Characterization . . . . . . . . . . . . . . . . . . . . . . . 39
3.8.1 Response to Organic Solvents . . . . . . . . . . . . . . . . . . 39
3.8.2 Spatial Correlation of Capacitance Sensors . . . . . . . . . . . 41
3.9 System Level Improvements . . . . . . . . . . . . . . . . . . . . . . . 43
v
3.9.1 Parallel Readout . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.9.2 Tunable Integration Time & Active Pixel Selections . . . . . . 44
3.10 Sensor Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.10.1 Temperature Sensitivity . . . . . . . . . . . . . . . . . . . . . 46
3.10.2 Power Supply Sensitivity . . . . . . . . . . . . . . . . . . . . . 47
3.10.3 Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4 Tracking Cell Metrics 51
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Device Packaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Tracking Cell Adhesion & Migration . . . . . . . . . . . . . . . . . . 55
4.5 Quantifying Cell Growth Rates . . . . . . . . . . . . . . . . . . . . . 58
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5 Real-time Imaging in Lab-on-CMOS Applications 63
5.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1 Visualizing Cells . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2.3 Alternative Imaging Systems . . . . . . . . . . . . . . . . . . . 67
5.2.4 Previous Approaches . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 This Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3.1 In-incubator Imaging . . . . . . . . . . . . . . . . . . . . . . . 70
5.3.2 Dual-Well Approach . . . . . . . . . . . . . . . . . . . . . . . 71
5.4 Live Cell Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4.1 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . 72
5.4.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . 73
5.5 Cell Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.5.2 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5.2.1 Pre-processing . . . . . . . . . . . . . . . . . . . . . 79
5.5.2.2 Cell Detection . . . . . . . . . . . . . . . . . . . . . 80
5.5.3 Results & Discussion . . . . . . . . . . . . . . . . . . . . . . . 82
5.5.3.1 Sensor Correlation . . . . . . . . . . . . . . . . . . . 82
5.5.3.2 Sensor Gain . . . . . . . . . . . . . . . . . . . . . . . 85
6 Applications, Part 1: Mitotic Activity Detection 87
6.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2.1 Mitotic Counts . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2.2 The Cell Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.3 Detecting Morphology Changes . . . . . . . . . . . . . . . . . . . . . 90
6.4 Mitosis detection algorithms . . . . . . . . . . . . . . . . . . . . . . . 91
6.4.1 Kaufman Adaptive Moving Average . . . . . . . . . . . . . . . 91
vi
6.4.2 Template Matching . . . . . . . . . . . . . . . . . . . . . . . . 92
6.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.5.1 Mitotic Activity Detection . . . . . . . . . . . . . . . . . . . . 93
6.5.2 Detection Algorithm Performance . . . . . . . . . . . . . . . . 96
7 Applications, Part 2: Drug Screening on Chip 98
7.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2 Cell-Based Drug Screening Assays . . . . . . . . . . . . . . . . . . . . 98
7.3 Cell Response to Chemotherapeutic Agents . . . . . . . . . . . . . . . 100
7.4 Cell Death Kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
8 Conclusions 107
8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
8.2 Comparison to State of the Art . . . . . . . . . . . . . . . . . . . . . 108
8.2.1 Cell-based Assays . . . . . . . . . . . . . . . . . . . . . . . . . 108
8.2.2 CMOS Capacitance Sensors . . . . . . . . . . . . . . . . . . . 109
8.2.3 Capacitance Resolution . . . . . . . . . . . . . . . . . . . . . . 111
8.2.4 System Complexity . . . . . . . . . . . . . . . . . . . . . . . . 113
8.2.5 Temporal Resolution and Hardware Interface . . . . . . . . . . 115
8.2.6 Imaging on CMOS ICs . . . . . . . . . . . . . . . . . . . . . . 118
8.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.3.1 Increased Spatial Density . . . . . . . . . . . . . . . . . . . . 118
8.3.2 Microfluidic Integration . . . . . . . . . . . . . . . . . . . . . 120
8.3.3 Closed-loop Feedback . . . . . . . . . . . . . . . . . . . . . . . 120
8.3.4 Sensitivity Improvement . . . . . . . . . . . . . . . . . . . . . 121
8.3.5 Surface Functionalization . . . . . . . . . . . . . . . . . . . . . 123
8.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.4.1 Drug Development - Electronic Multi-well Plate . . . . . . . . 123
8.4.2 Point-of-care Diagnostics and Personalized Medicine . . . . . . 126
8.5 Publication Record . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.5.1 Refereed Journal Articles . . . . . . . . . . . . . . . . . . . . . 126
8.5.2 Journal Articles in Preparation . . . . . . . . . . . . . . . . . 127
8.5.3 Refereed Conference Proceedings . . . . . . . . . . . . . . . . 127
A Appendix 129
A.1 Sensor Design Details . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
A.2 Sample Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Bibliography 137
vii
List of Tables
4.1 Cell Growth Rate Estimation . . . . . . . . . . . . . . . . . . . . . . 59
8.1 Biosensors from State of Art . . . . . . . . . . . . . . . . . . . . . . . 117
List of Figures
1.1 Conceptual diagram of sensor system. . . . . . . . . . . . . . . . . . . 2
2.1 Schematic of the charge sharing measurement system. [1] . . . . . . . 7
2.2 Schematic of the CBCM test structure. [2] . . . . . . . . . . . . . . . 9
2.3 Schematic of frequency based capacitance readout circuit. [3] . . . . . 10
2.4 Cell dielectric layer formation on a solid substrate. Left: Cells in
suspension first drift downwards to settle on the surface. They begin
adhering to the surface through various mechanisms involving cell
adhesion molecules [4]. The cells then anchor and spread outward
as they proliferate. Right: Microscope images of cells cultured on
the biosensor surface, during the settling, adhering, and proliferating
phases (top to bottom). . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Schematic of the ECIS measurement system. [5] . . . . . . . . . . . . 14
3.1 A model of the capacitances at the cell-substrate interface. . . . . . . 21
3.2 A model of the capacitances at the cell-substrate interface with in-
terdigitated electrodes. . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Sensed capacitance as a function of the relative permittivity of the
environment over the input electrodes, for a) the analytical model, b)
the FEM simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 a) Circuit diagram of capacitance sensor pixe, b) System block diagram 27
3.5 Output frequency as a function of sensed capacitive load. Results for
typical expected capacitances are shown in red. The dotted line shows
Monte Carlo simulation results with the inset showing a histogram
for a sensed capacitance of 10 fF. . . . . . . . . . . . . . . . . . . . . 29
3.6 Output frequency for the temperature and capacitance sensors as a
function of temperature. . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.7 Photomicrograph of a) Fabricated die (3 ? 3 mm2), b) sensing array,
and c) sensing electrodes (30 ? 30 ?m2). . . . . . . . . . . . . . . . . 35
3.8 a) Exeprimental setup to characterize sensor as a distance sensor. b)
Sensor output frequency as a metal probe is moved away from the
surface of the chip. Solid line shows a regressive fit of the measured
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.9 Sensor characterization curves generated by mapping electrode probe
distance to modeled capacitance value. The curves from left to right
indicate the best, average, and worst case sensitivity scenarios. . . . . 38
viii
3.10 Sensor response to addition of different fluids. (a) Transient plot of
response for a single pixel. (b) Mean response and standard deviation
across 16 pixels. (c) Mean response as a function of relative dielectric
permittivity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.11 Sensor response due to application of difference concentrations of iso-
propyl alcohol. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.12 Mapping sensor response to external probe interacting with chip sur-
face. a) 16 pixel responses plotted over time as the probe is brought
down and lifted off. b) Shaded plot mapping responses to physical
pixel locations. Each array image corresponds to time points indi-
cated by the dashed lines in (a). . . . . . . . . . . . . . . . . . . . . . 42
3.13 a) Expanded system to include column-wise readout. b) Die photo
of the 2nd generation chip . . . . . . . . . . . . . . . . . . . . . . . . 43
3.14 a) System block diagram of chip with programmable integration time,
pixel activation and programmable I2C address. b) Die photo of the
3rd generation chip . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.15 Sensitivity and compensation for temperature (top) and power supply
(bottom) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.16 Long-term sensor response and noise when exposed to cell media at
37 ?C. a) Output frequency vs. time for each pixel (gray) and their
mean (black). b) Histogram of the mean frequency. . . . . . . . . . . 50
4.1 Cross-sectional views of chips packaged in a) an epoxy handle wafer,
and b) a passivated DIP carrier . . . . . . . . . . . . . . . . . . . . . 52
4.2 a) Photographs of the capacitance sensor chip packaged using the
three different methods (from left to right: ceramic carrier, epoxy
carrier, LTCC carrier). b) The test setup. Left: image of a packaged
chip on a daughterboard connected to the readout microcontroller
using an Ethernet cable. Right: image of the daughterboard inside
the incubator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 a) Response curves of the chip as CP70 cells adhere and proliferate
across the sensor array. The two vertical black lines at t = 21 hours
and t = 44 hours indicate times when data recording was paused for
imaging. Right panel shows image of surface of chip after b) 21 hours,
and c) 44 hours of incubation of CP70 cells. . . . . . . . . . . . . . . 56
4.4 Results from two cell experiments performed after cleaning and reusing
a device. Lines show output from sensors with (red) and without
(black) cell coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.5 Cumulative number of sensors that show an increase in capacitance
as a function of time, for each of the four experiments described in
Section 4.3. The large plateau after 22 hours in (b) is due to solution
evaporation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.6 Plot of the measured sensor response as a function of estimated cell
coverage of pixels. a) All experiments separately, b) Mean value.
Linear fit shows a sensitivity of 100 aF/cell. . . . . . . . . . . . . . . 60
ix
4.7 Plot of the measured change in capacitance as a function of change
in cell coverage (Device 1, CP70 cells) . . . . . . . . . . . . . . . . . . 61
5.1 Photomicrographs of cells on CMOS chips. Left: Images of sensory
neurons taken with a) standard objective lens, b) immersion lens.
Image from [6]. Right: Images of ovarian cancer cells, with media
evaporation resulting in loss of focus after 5 hours. . . . . . . . . . . . 66
5.2 Schematic of a borescope-based imaging system. Placement of the
lens within the cell solution can help to reduce effects of light refraction. 68
5.3 a) Schematic of dual-well package. b) Photograph of packaged chip
inside an incubator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4 Correlation of capacitance measurements and time-lapse imaging of
cells on the chip during cell-substrate binding. a) Images taken of
four sensors at different time points. b) Corresponding sensor re-
sponses showing the measured capacitance changes over time. The
letter markers indicate the time points corresponding to each frame.
Video is available as supplementary material. . . . . . . . . . . . . . . 74
5.5 Correlation of capacitance measurements and time-lapse imaging of
cells on the chip during cell-substrate binding. a) Images taken of
four sensors at different time points. b) Corresponding sensor re-
sponses showing the measured capacitance changes over time. The
letter markers indicate the time points corresponding to each frame.
Video is available as supplementary material. . . . . . . . . . . . . . . 76
5.6 Images of intermediate steps in cell segmentation algorithm. a) Origi-
nal image of cells over sensor electrode. b, c) Result of registering and
subtraction of bare reference electrode for dark (b) and bright (c) em-
phasis. d) Output of two-level adaptive thresholding with orange and
red regions corresponding to bright and dark levels, respectively. e)
Mask after morphological filtering. f) Detected cell coverage regions
and electrode area, marked in red and orange outlines, respectively. . 80
5.7 Correlation of measured capacitance and estimated cell coverage. a)
Capacitance and coverage plotted as a function of time for two exem-
plar datasets. b) Scatter plots of capacitance vs. estimated coverage,
with linear regression line, for the data in (a). c) Histogram of the
difference between fit line and measured capacitance values for all
datasets (N = 3196). d) Derived sensor gain vs. correlation coeffi-
cient. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.1 Example images of nuclei that are challenging to detect (figure taken
directly from [7]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
x
6.2 a) Top: Images taken of a sensor at different time points. Bot-
tom: Corresponding sensor response showing the measured capaci-
tance changes over time. The highlighted portions indicate division
events marked by the detection algorithm. Inset: Distribution of the
time taken from start of division event to reach minimum capacitance
(Tdrop) and time taken to recover (Trise). N = 41 detected events. b)
Examples of detection mitosis events from one experiment. Capaci-
tance traces have been offset for plotting purposes. Video is available
as supplementary material. . . . . . . . . . . . . . . . . . . . . . . . . 94
6.3 Receiver operating characteristic curve for cell detection algorithm.
The two lines correspond to varied minimum capacitance changes and
durations for mitotic events. . . . . . . . . . . . . . . . . . . . . . . . 97
7.1 Top: Microscope images of the two sensors during a drug adminis-
tration experiment. Bottom: Corresponding sensor measured capac-
itance changes over time. The cisplatin was added at t = 53.5 hours,
indicated by the red line and arrow. The labels on the images corre-
spond to the time points marked by the dashed vertical lines. Video
is available as supplementary material. . . . . . . . . . . . . . . . . . 101
7.2 Mean capacitance sensor responses for various concentrations of ap-
plied drug. Each chart shows the response for two cell types, A2780
(drug-sensitive), and CP70 (drug-resistive). The dashed lines indi-
cate when the drug was applied to the cell culture. In the control
experiment both cell lines were viable through the experiment. Addi-
tion of 1 ?M cisplatin showed an immediate effect on the capacitive
coupling of A2780 cells while CP70 cells remained stable. Higher
concentrations began to induce a response in CP70 cells as well. . . . 102
7.3 Measured difference in average capacitance from immediately after
the drug was administered to 24 hours afterwards. . . . . . . . . . . . 103
7.4 Normalized capacitance values vs. time for experiments where cells
were exposed to drug concentrations of a) 25 ?M and b) 100 ?M .
The bottom panel shows the mean extracted a) time delay and d)
decay time constant for the set of experiments. (? ? p < 0.01) . . . . . 105
8.1 Alamar Blue assay procedure. Figure from [8] . . . . . . . . . . . . . 108
8.2 Example plots of data from CMOS-based capacitance sensors used for
in vitro experiments. Images taken from a) Prakash et al. 2008 [9],
b) Couniot et al. 2015 [10], c) Nabovati et al. 2018 [11], d) Nabovati
et al. 2017 [12], e) Laborde et al. 2015 [13] . . . . . . . . . . . . . . . 110
8.3 Block diagram for a feedback control loop to perform automated drug
dispensement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.4 FEM simulation results showing the effect of passivation layer thin-
ning on sensed capacitance. The electrode and cell structure pre-
sented in Section 3.2 was simulated as cell radius was increased with
a passivation thickness of a) 1.6 ?m, and (b) 0.8 ?m. . . . . . . . . . 122
xi
8.5 Schematic of an ?electronic multi-well plate?. Each well contains
a chip with its sensing area exposed to the cell culture. Electronic
connections are made underneath the well plate across columns and
rows on a daughterboard. The final routing is wired to a connector
that can then be read using a microcontroller. . . . . . . . . . . . . . 125
A.1 Circuit schematic of on-chip bias generator. . . . . . . . . . . . . . . 129
A.2 Finite-state machine diagram of on-chip logic. . . . . . . . . . . . . . 130
A.3 Simulation results of logic for readout of one pixel (accumulation time
shortened to 1 ?s for simulation purposes. The figure depicts a a) full
view, and b) zoomed-in view. . . . . . . . . . . . . . . . . . . . . . . 131
A.4 a) Response curves of the chip as CP70 cells adhere and proliferate
across the sensor array. The first two vertical black lines at t =
25 hours and t = 47 hours indicate times when data recording was
paused for imaging. Right panel shows image of surface of chip after
b) 25 hours, and c) 47 hours. . . . . . . . . . . . . . . . . . . . . . . . 132
A.5 a) Response curves of the chip as A2780 cells adhere and proliferate
across the sensor array. The first two vertical black lines at t = 28
hours and t = 49.5 hours indicate times when data recording was
paused for imaging. Right panel shows image of surface of chip after
b) 28 hours, and c) 49.5 hours. . . . . . . . . . . . . . . . . . . . . . . 133
A.6 a) Response curves of the chip as CP70 cells adhere and proliferate
across the sensor array. The first two vertical black lines at t =
28 hours and t = 48 hours indicate times when data recording was
paused for imaging. Right panel shows image of surface of chip after
b) 28 hours, and c) 48 hours. . . . . . . . . . . . . . . . . . . . . . . . 134
A.7 Results of experiment on Device 1 with A2780 cells. Microscope
image of chip surface after a) 21.5 hours and b) 46 hours of incubation.
c) Response curves where sensors with and without cell coverage are
grouped together. Two sensors observed to have no coverage at the
two time intervals (2,4) and (4,3) are shown in dotted red, with their
mean as a solid line. The other sensors are shown in dotted black,
with mean as a solid line. . . . . . . . . . . . . . . . . . . . . . . . . . 135
A.8 Results of experiment with A2780 cells. Microscope image of chip
surface after a) 27 hours and b) 70 hours of incubation. c) Response
curves of sensors with cell coverage in gray, with the mean as a solid
black line. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
xii
List of Abbreviations
ADC analog-to-digital converter
CAN controller area network
CBCM charge-based capacitance measurement
CMOS complementary metal-oxide semiconductor
CS charge sharing
CTF capacitance to frequency
DI deionized
ECIS electric cell-substrate impedance sensing
FEM finite element method
FPGA field-programmable gate array
FPR false positive rate
I2C inter-integrated circuit
IBIS Integrated Biomorphic Information Systems
IPA isopropyl alcohol
KAMA Kaufman adaptive moving average
LoC lab-on-chip
LoCMOS lab-on-CMOS
LOD limit of detection
LTCC low-temperature co-fired ceramic
MAI moving average indicator
PCB printed circuit board
ROC receiver operating characteristic
RPMI Roswell Park Memorial Institute
SPI serial peripheral interface
SSD sum of squared difference
TPM template matching
TPR true positive rate
xiii
VCO voltage-controlled oscillator
xiv
Chapter 1
Introduction
1.1 Lab-on-CMOS Systems
Microelectronic biosensors have the potential to provide low-cost, adaptable,
and portable devices for applications in fields such as biology, environmental science,
and medicine. In particular, a ubiquitous experimental challenge is to monitor the
proliferation of cells over time to determine their response to stimuli or environ-
mental conditions. Monitoring cell growth and migration is an important tool for
optimizing cell culture conditions, testing drugs, and screening materials for bio-
compatibility and toxicity [14]. Conventional approaches involve the use of dyes
and markers that can potentially introduce side-effects into the cell culture [15] and
often function as end-point assays, eliminating the opportunity to track fast changes
and to determine temporal correlation between measurements. Therefore it is im-
portant to develop new techniques to perform these measurements. Lab-on-chip
devices have shown promise in changing the way that biological samples are ana-
lyzed. They allow for smaller sample volumes, reduced cost, faster reaction kinetics,
and higher throughput than conventional approaches [16, 17]. Another important
advantage of lab-on-chip systems is their potential for automated and unsupervised
data collection. In most instances, however, lab-on-chip devices comprise of a pas-
sive sensing layer that requires additional external instrumentation for readout and
1
Capacitance sensor for on-chip cell Real-time imaging of lab- Proof-of-concept drug 
monitoring on-CMOS chips screening tool
CMOS IC
Imager
Serial 
Interface
CIN
Control 
Logic
MUX
VB
VO IC Chip
EN
Row/Col. 
Decoders Chip Carrier
Time
Figure 1.1: Conceptual diagram of sensor system.
signal processing [6, 16].
Recently, complementary metal-oxide semiconductor (CMOS) technology has
become an attractive platform for lab-on-chip systems since it facilitates the incor-
poration of sensing and actuation in intimate contact with readout and electronics,
producing systems with smaller size, higher sensing density, and higher sensing
fidelity. CMOS technology offers the advantages of having low static power require-
ments and good noise immunity [16], as well as being widely available and facilitating
integration with commercial electronics. This research describes the development of
a fully integrated lab-on-CMOS (LoCMOS) capacitance biosensor for on-chip cell
sensing applications. The specific focus of the work is on developing a low-cost
and low-complexity portable system that can be used for live cell experiments, in-
cluding drug screening assays. The CMOS sensor chip aims to characterize the
adherence of biological cells to a substrate, which is an important indicator of cell
viability, through high-resolution surface capacitance measurements. This sensing
paradigm relies on the biophysical morphology of adherent cells. Healthy cells with
2
|?C| (aF)
well-formed membranes create small changes in the dielectric properties of the area
around the electrodes [18, 19]. Additionally these cells bind tightly to their sub-
strates in comparison to dead or compromised cells [20]. These properties provide
an avenue for monitoring cell health through an integrated LoCMOS system [9]. An
imaging platform for CMOS chips is then presented that allows for simultaneous
data recording and image acquisition to provide ground-truth data for sensor val-
idation. Finally the LoCMOS system is used as a proof-of-concept drug screening
tool. An overview of the goals of this research is shown in Fig. 1.1.
1.2 Research Contributions
The primary contributions of this work are highlighted in the points below:
? Design of CMOS capacitance sensors for on-chip cell monitoring.
A low-cost, low-complexity CMOS sensor was developed that provides high-
resolution capacitance measurements. The chip is designed with on-chip signal
digitization and serial communication, which facilitate sensor packaging for
experiments, and allows for readout with any commercial microcontroller. No
large laboratory equipment or electrical hardware is required for operation,
enabling future extensions to point-of-care (PoC) devices.
? Real-time imaging of lab-on-CMOS chips. An imaging platform was
developed for simultaneously making recordings using a CMOS sensor and
taking images of the sensor surface. This allowed for validation of sensor
measurements by providing ground-truth images for cell sensing applications.
3
The imaging does not interfere with cell viability and maintains image quality
over several days of experiments.
? Drug screening on chip. Proof-of-concept experiments are performed to
characterize the potency of chemotherapeutic drugs on cells using the lab-on-
CMOS capacitance sensor. Results showed the ability to discriminate drug-
sensitive and resistive cell lines using capacitance measurements. The high
temporal-resolution data provided by the chip allowed quantification of cell
death kinetics.
1.3 Thesis Organization
Chapter 2 provides background information on cell-substrate adhesion, and
techniques to monitor cell viability. An overview of CMOS-based capacitance sen-
sors is also presented. Chapter 3 introduces the capacitance sensor developed
in this work. The front-end design is discussed along with a system-level descrip-
tion of the chip, and simulation results. The fabricated chip is characterized on the
benchtop to extract performance metrics. Chapter 4 presents results of in vitro ex-
periments using cancer cell lines that are grown on the sensor. The chapter discusses
packaging methods, the experimental protocol, results, and presents quantification
metrics to track population cell growth. Chapter 5 introduces the real-time imag-
ing platform. Background material on visualizing cells is presented. The hardware
implementation is described and results of simultaneous imaging and capacitance
recording in in vitro experiments are shown. Images are used to validate the capac-
4
itance sensor measurements and show the ability to detect single-cell morphology
changes. Finally, preliminary work on image processing to perform automatic cell
detection and coverage estimation is presented. Results are correlated with the data
recordings from the LoCMOS sensor. Chapters 6 and 7 present work towards using
the chip as a tool for drug development. Chapter 6 shows results of using the
sensor to detect cell mitosis activity, is an important prognostic factor and indicator
of metastatic risk in cancer patients. Chapter 7 presents results of drug screening
assays done on chip where drug-sensitive and drug-resistive cell types are exposed to
a cancer drug. Results show the ability to discriminate cell lines using capacitance
sensor measurements. The data is then used to extract cell death kinetics.
5
Chapter 2
Capacitance Sensing on CMOS
2.1 Sensing Approaches
Many types of CMOS-based capacitance sensors have been developed and used
in various applications. The fundamental sensing mechanism is to detect changes in
the dielectric properties of a material with respect to a baseline such as air or liquid
solution. Common applications of these types of sensors include detecting proximity,
humidity, and force detection. One of more ubiquitous applications in modern times
are their use in human interface devices such as trackpads, smartphone screens, and
non-mechanical buttons. A number of different transduction methods have been
employed in CMOS technology for use in capacitance measurements. Two specific
mechanisms that have been developed for biological applications are charge-based
and time/frequency-based sensors.
2.1.1 Charge based capacitance measurements
Charge based capacitance sensors have been used in many prior applications,
initially starting with characterizing interconnect capacitances [21, 2] and fingerprint
sensing [1, 22]. This transduction mechanism has since been adapted for use in
numerous biological cell applications [9, 23, 24, 25, 10].
6
2.1.1.1 Fingerprint Sensor
Lee et al. introduced a fingerprint sensor chip based on an array of capacitance
sensors that use the charge sharing (CS) sensor scheme [1]. Fig. 2.1 illustrates this
technique. The input analyte (finger) acts as the upper electrode of the capacitor,
and the metal plate on the chip acts as the other electrode. The capacitance CS is
related to the capacitance being sensed and will decrease as the distance between
the finger and the passivation layer increases. Therefore CS will be at its maximum
value when a ridge on the finger contacts the passivation layer. Capacitances Cp1
and Cp2 represent parasitic capacitances on nodes N1 and N2.
Sensing of the capacitance works in two phases: precharge and evaluation. In
the first precharge phase, S1 and S3 are on while S2 is off, causing N1 and N2 to
finger 
CS air 
passivation 
S1 S2 S3 
N1 N2 
V1 V2 
Cp1 Cp2 
Figure 2.1: Schematic of the charge sharing measurement system. [1]
7
charge to known voltages V1 and V2, respectively. In the next stage the switches
are flipped. This causes the charges on the two nodes to re-distribute, giving a final
output voltage of:
Cp1V1 + Cp2V2 + CSV1
VO = VN1 = VN2 = (2.1)
Cp1 + Cp2 + CS
The output voltage is thus related to the input sensed capacitance, and it
drops off rapidly as the distance between finger and sensor surface increases [1].
2.1.1.2 Metal Interconnect Characterization
Another charge-based sensing scheme is charge-based capacitance measure-
ment (CBCM) which was introduced by Chen et al. to analyze parasitic capac-
itances of overlaid metal interconnents on CMOS chips [2]. Fig. 2.2 shows the
CBCM-based test structure used for this characterization. The structure consisted
of identical PMOS and NMOS transistors that were controlled by non-overlapping
signals V1 and V2. The capacitance being sensed (CS) was attached to one pair of
transistors. In Fig. 2.2, this amounts to the addition of a grounded metal 2 line
that overlaps the existing metal 1 line connected to the right-pair?s middle node.
The control signals V1 and V2 alternately charge and discharge the intercon-
nect capacitance at a frequency f . The capacitance can be derived by finding the
difference between the DC currents I and I ? (measured externally using an amme-
ter):
Inet = I
? ? I = CSVDDf (2.2)
8
VDD-NoCap VDD-Cap 
I? I 
V  T = 1/f 2 CS VDD 
V1 
0 
V1 VDD 
V2 
0 
Figure 2.2: Schematic of the CBCM test structure. [2]
The actual capacitance is computed by finding the slope of the Inet-VDD curve
and dividing by the frequency [2]. In the ideal case, both sides of the CBCM struc-
ture will be identical, having the same drain overlap and drain junction capacitances.
However, any mismatch between the two sides will limit the minimum achievable
resolution.
2.1.2 Frequency-based Capacitance Measurements
The capacitance to frequency (CTF) sensing technique uses capacitive loading
to modulate the frequency of oscillating circuits. The oscillating element in these cir-
cuits generally consist of a capacitor along with a resistor or inductor. For integrated
CMOS sensors, the capacitor is replaced directly with the input sensing electrodes,
while the resistor (for an RC element) simply consists of parasitics. Charging and
discharging of the element is generally performed using current, resulting in an out-
?
put frequency that is proportional to 1/RC or 1/ LC. As an example, Lu et al.
9
? V 
V 
Time 
Time 
+ Delay 
V  
C  V  ? Stage oOX SW
RCT 
C C  CP elec
RS 
Figure 2.3: Schematic of frequency based capacitance readout circuit. [3]
used a CTF technique to detect the presence of the neurotransmitter dopamine, the
binding of which changed the charge profile at the sensor interface, inducing changes
in capacitance [3]. Fig. 2.3 shows the CTF-based readout circuit.
The input capacitance (Celec) lies in parallel with parasitics that include in-
terface capacitive impedance (CCP ), oxide capacitance (COX), charge transfer resis-
tance (RCT ), and solution resistance (RS). RCT and RS are assumed to be negligible
and so the overall impedance is purely capacitive. A tuneable bank of current sources
that range from 1 ?A to 16 ?A is used to charge Celec, producing a ramping signal.
When this signal reaches a certain switching voltage (VSW ), the comparator out-
put goes high and after a delay stage, causes the capacitor charge to reset through
an NMOS transistor. This feedback loop sustains a continual generation of output
pulses at a frequency:
[ ]?1
CSVSW
f = + ?d (2.3)
IB
10
where IB, ?d, and CS are the charging current, time delay of the output stages, and
overall input capacitances, respectively.
2.2 Cell Adhesion & and Viability sensing
2.2.1 Cell-Substrate Coupling
Adherent cells modify the dielectric properties at the cell-substrate interface
as they settle onto a substrate, adhere, proliferate, and eventually lift off due to
cell death, forced detachment, or other morphological changes. Fig. 2.4 shows a
diagram of three stages of the cell adhesion process. When cells are first placed in
suspension they drift downwards and make an initial attachment with the substrate
(top panel). They then begin attaching to the surface through a variety of cell
adhesion mechanisms (middle panel) [4]. The cells then anchor and flatten them-
selves, spreading outward over a larger area until they reach their maximum spread
(bottom panel) [26]. Healthy cells will continue to bind tightly to the substrate and
spread out as they grow and proliferate, while compromised cells contract and may
even lift off from the surface entirely. These changes in cell morphology modulate
the dielectric properties of the cell-substrate interface, which can be measured as
changes in surface capacitance [9].
2.2.2 Non-Electronic Techniques
Numerous non-electronic methods for quantifying cell adhesion have been ex-
plored [26]. Traction force microscopy is a technique that involves culturing cells
11
Cell Adhesion
cells settling cell media
substrate
cells adhering
cells proliferating
50 ?m
Figure 2.4: Cell dielectric layer formation on a solid substrate. Left: Cells in suspension
5 Integrated Biomorphic Information Systems Lab
first drift downwards to settle on the surface. They begin adhering to the surface through
various mechanisms involving cell adhesion molecules [4]. The cells then anchor and spread
outward as they proliferate. Right: Microscope images of cells cultured on the biosensor
surface, during the settling, adhering, and proliferating phases (top to bottom).
onto a polyacrylamide gel that has been embedded with fluorescent beads. When
cells adhere they generate traction forces that can be quantified by monitoring the
fluorescent beads [27]. Another technique involves the exposure of cells to shear fluid
flow through microfluidic channels [28]. The strength of cell adhesion is related to
the balance between the adhesion strength and the shear fluid forces. Another me-
chanical technique of quantifying adhesion is to use a micropipette to apply a suction
force to the top of a cell. The minimum amount of force required to pull the cell of
the substrate is a measure of its adhesion strength [29].
Traditional methods of assessing cell viability involve direct visual counting of
healthy cells or indirect measurement of a separate biomarker. One method involves
incubating cells with a reagent (such as Alamar Blue) that develops a visible color
through the metabolic activity of active cells. Measurement of the color development
is performed using spectrophotometry and is then an analogue for cell viability
[30, 31]. Another method of assessing cell viability is to analyze the health of cell
12
membranes. When a dye, such as trypan blue, is applied, healthy cells with intact
membranes will exclude the dye, whereas dead cells will be stained. The number
of cells that have not been stained can then be counted under a microscope to get
an indication of viability. These optical techniques typically require microscopes
and optical filters for measurement. Methods that require the addition of a dye or
fluorescent agent can compromise viability and lead to cell death [32]. Additionally
these approaches require sampling of the analyte, which might not be feasible for
small culture volumes. Therefore a label-free detection method would be beneficial.
Furthermore, many traditional characterization methods are end-point assays which
provide a measure of adhesion only at the end of an experimental procedure. Many
parallel experiments are required to obtain a temporal sequence of discrete sample
points, with temporal accuracy limited by the variability across samples and the
sampling technique. Much richer information about the dynamics of cell growth
would be available from a real-time measurement method.
A number of physical transduction mechanisms have been introduced for mon-
itoring biological cells. Impedance spectroscopy, quartz crystal microbalances, and
surface acoustic waves [33, 34, 35], commonly used for characterization of mate-
rial properties, have been used to study cells; however they generally require large
external equipment for readout. Ion-sensitive FETs have also been used to detect
bacterial activity [36], but they are generally used to measure analogues of cell ac-
tivity (such as pH changes), rather than viability directly. Amperometry has also
been used in cell detection [37], however post-fabrication of electrodes on CMOS
chips are required.
13
2.2.3 Electric Cell-Substrate Impedance Sensing
One of the most popular and well-established electronic methods of cell sensing
is electric cell-substrate impedance sensing (ECIS), first introduced by Giaever and
Keese [38, 39]. This measurement technique monitors the impedance across a pair
of sense and counter electrodes. In general, a 1 V amplitude AC signal is applied to
the electrodes through a 1 M? resistor at a given frequency (nominally 4 kHz). A
separate lock-in amplifier is used to detect the phase and magnitude of the resulting
signal. A schematic of the readout system is shown in Fig. 2.5.
The detected signal is fitted to an equivalent resistor and capacitor model to
extract barrier resistance, membrane capacitance, and current flow parameter values
[5]. The resistance and capacitance of the electrodes (typically gold) are assumed
not to change, so any measured changes in resistance and capacitance values is due
to the cells altering the current flow.
ECIS systems are commercialized and have been used in several studies. Gi-
Sense Counter 
Electrode Electrode 
1 M? 
Data 
Lock-in 
Acquisition & 
Amplifier 
Processing 
Figure 2.5: Schematic of the ECIS measurement system. [5]
14
aever and Keese initially performed studies on cell attachment, motion, and effects
of metabolite application [40, 5, 38]. Other research has also been performed using
commercial ECIS systems. Balasubramanian et al. studied changes in cell morphol-
ogy induced by integrin binding peptides [33]. Opp et al. monitored cytotoxicity of
human endotheliel cells to different dosages of metabolies [41]. Allen-Gipson et al.
observed wound healing characteristics of bronchial epithelial cells in the presence
of cigarette smoke extract [42].
2.2.4 Integrated Sensing on CMOS
CMOS-based capacitance sensors have been developed for a number of cell
monitoring applications. These chips implement arrays of sensing electrodes that
detect capacitive changes as cells adhere to and grow across the surface of the chip.
The electrodes are generally insulated from the cell culture environment and so limit
undesirable electrochemistry that may otherwise occur.
Both charge sharing and charge-based capacitance measurement sensing have
been used in LoCMOS systems for live cell sensing experiments. Prakash et al.
used a CS-based chip to track adhesion and proliferation of smooth muscle and
breast cancer cells [43, 9]. Ghafar-Zadeh et al. used a CBCM structure to measure
growth of bacteria in medium [44, 25]. This system included a ?? ADC to convert
the output voltage to a serial stream of bits that is read off-chip using a field-
programmable gate array (FPGA). Prakash et al. extended the single-ended CBCM
structure to a fully differential version with rail-to-rail output [45, 46]. It showed a
15
measured resolution of 15 fF based on benchtop experiments.
As described previously, any mismatch in the CBCM structures can limit the
resolution of the sensors. Additionally, if current subtraction is done on-chip it is
possible for the parasitic mismatches to drive the output signal into saturation at
the zero-level baseline [23, 47]. One method to combat this is to use a bank of
reference currents in the CBCM structure to do offset cancellation [24, 12]. Alterna-
tively, Prakash et al. extended the differential CBCM design to include floating-gate
trimming circuits to perform the mismatch and offset cancellation [23].
For the CTF paradigm, Cuoniot et al. used a five-stage ring oscillator loaded
with five separate large interdigitated electrodes (total area 100 ? 100 or 200 ? 200
?m2) to detect bacteria in saline buffer solution [48]. Table 8.1 presents a summary
of existing work.
2.2.5 Discussion of Integrated CMOS Approaches
Mismatch & Offset Capacitance. Single-ended CS or CBCM systems
have high sensitivity to capacitive changes but can be limited in resolution due
to parasitics. Additionally, the permittivity of the immediate surroundings of the
electrode can create a large offset capacitance that can reduce the effective dynamic
range of the sensor. The differential CBCM systems described previously were
shown to compensate for parasitics, to increase dynamic range, and to give high
resolution readings, especially when post-fabrication compensation can be completed
(e.g. using floating gate trimming [23]). However, these systems still require a
16
complex readout system and initial programming stage. Capacitance to frequency
based transduction systems have the benefit of being inherently resistant to large
offset capcitances and parasitic mismatch. This is because there is technically little
limit to the range of output frequencies at which the oscillators run. If the integration
time is long enough, even very slow signals (large capacitances) can be resolved.
Single-ended CTF sensors are still subject to the environmental permittivity, which
can cause a large baseline frequency shift. This can be removed in post processing,
but a differential CTF structure could negate this offset. Damilano et al. [49]
developed a CTF readout circuit for tactile skin sensors that uses a second electrode
to generate a reference frequency that was used to subtract out offsets, creating a
?pseudo-differential? setup. Simulation results showed a 10 fF sensitivity per output
LSB, with a pixel size of 22? 79 ?m2.
Analog to Digital Conversion. Another important point of consideration
for a lab-on-CMOS platform is digitization of the sensor output, since analysis of
recorded data will almost exclusively be done digitally. For traditional CS or CBCM
sensors, the output signal is an analog voltage that must be sent off-chip to an ex-
ternal ADC. However, this provides a path for the addition of noise into the analog
signals and so careful shielding is necessary [50]. In a cell culture environment this
may be difficult. As mentioned previously, CBCM sensors with on-chip ?? ADCs
have been implemented [25, 12]. In these systems an external FPGA is used to
provide the switching signals for the CBCM core circuitry, to decode the ?? bit-
stream, and to create the array readout controls. This setup has been shown to
work in experiments with polystyrene beads and live cells. However, the use of an
17
FPGA, and the complexity of the signal transduction and control methods increases
the required chip readout requirements and may have adverse effects on total sys-
tem power consumption. This limits future sensor development for use in portable
applications. Frequency based systems are inherently suited for digitization because
the output signal is digital, with capacitance information encoded in its oscillation
frequency. A straightforward method of converting the frequency information into
a digital format is to use it to drive a counter [51, 49].
System Readout. It is important to design CMOS chips with packaging
considerations in mind. This is especially important for cell sensing applications,
since adequate passivation of the exposed electrical traces is required in wet environ-
ments. Possible future incorporation of microfludics is also an important factor for
consideration [50]. With this in mind, minimizing the number of I/O pads used by
the chip is a relevant goal since this will limit the number of possible failure points
during packaging. The number of bondpads required depends on how many external
bias voltages are required, the number control signals needed for chip control, and
how data from an array of sensors is read. On-chip bias circuitry can be used to
limit the need for external pins. The control and readout pins generally depend on
how many sensors are implemented on chip and the transduction method. In order
to reduce the number of pins required, parallel or sequential readout can be used.
CBCM-based systems have been used with a parallel readout format where sensors
in an array were read in a column-wise fashion [23, 12]. External row selection
signals were used to cycle row-by-row through all columns of the array, and data
was read through the Ncol output pins, where Ncol is the number of columns in the
18
array.
A sequential readout system cycles through each sensor in the array. This
means a single set of readout lines can be used for the entire array, at the cost of a
slower total readout speed. Readout of digital data can be further streamlined by
reading data out serially. To this end one could leverage ubiquitous serial commu-
nication protocols such as serial peripheral interface (SPI), inter-integrated circuit
(I2C) buses, or controller area network (CAN) buses. These systems generally have
two signal lines (clock and data) and can be interfaced with most modern microcon-
trollers. Another benefit is the ability to have multiple devices on the same readout
bus, thereby allowing for multiple chips to be interfaced together.
2.2.6 This Work
The standard methods of monitoring cell adhesion and viability outlined in
Section 2.2.2 have been used for a number of applications. However, they use special-
ized methods that often require the use of large laboratory-based equipment. These
methods only provide data at discrete time points, so higher throughput monitoring
would be labor-intensive. Additionally they generally involve physical removal of
samples for each observation, which might not be feasible for small sample volumes.
The ECIS technique outlined in Section 2.2.3 has shown success in numerous studies
varying from cell proliferation and toxicity to cell migration and wound-healing [52].
However, this technique still requires the use of large external equipment. Addition-
ally, the electrodes used in the system are exposed directly to the analyte and so
19
may be vulnerable to electrochemical effects in the long term. Another important
consideration is the sensitivity of the measurements to additive noise since signals at
the electrodes are routed to the measurement equipment through wires that may not
be minimally short. The integrated lab-on-CMOS approaches, on the other hand,
have the advantage of incorporating most of the sensing circuitry on chip along with
the sensing electrodes. This helps in reducing the need for specialized equipment.
Additionally, digitization can be performed on chip, helping to make the system
more robust to noise. Continuous recording can also be achieved, creating a system
for long-term label-free cell monitoring.
Here we develop an integrated CTF sensor that provides high-resolution mea-
surements with on-chip digitization and a readout system that minimizes I/O re-
quirements. The system provides unsupervised recording of measurements using
only low-cost equipment (< $50) and allows for high-throughput assays through the
use of parallel devices. Within the possible capacitance sensing paradigms outlined
in Section 2.2.4, the capacitance to frequency approach has numerous advantages.
As discussed in Section 2.2.5, the transduction mechanism can inherently manage
large variations in capacitance caused by changes in the permittivity of the cell me-
dia/environment. Additionally it offers simple data digitization with no additional
circuitry required. A high capacitance resolution can be achieved, but at the cost of
a long integration time (slow sampling rate). However for applications of cell sens-
ing, a low sampling rate is adequate since timescales of interest are on the order of
several minutes to hours. Therefore this work proposes to implement a CTF sensing
approach to cell based capacitance sensing.
20
Chapter 3
Capacitance Sensor Design
3.1 Cell Capacitance Model
As introduced in Section 2.2.1, adherent cells modify the dielectric properties
directly above the substrate interface. This creates a change in capacitive coupling
that is then sensed through metal electrodes that are fabricated underneath the
substrate on the CMOS IC. A schematic of the cell-substrate interface is shown in
Fig. 3.1. While the cell media and cell itself are conductive, the effect of parasitic
conductance/resistance is neglected because no current flow is expected. The pas-
sivation layer is assumed to be a very good insulating layer and so only capacitive
effects are considered [53, 9].
CIN
Passivation CPSV CPSV
Inter-layer
dielectric CP-fr V C
C
MM V P-fr1 2
CP-pp CP-pp
Substrate
Figure 3.1: A model of the capacitances at the cell-substrate interface.
21
The CMOS IC consists of the silicon substrate, layers of dielectric (silicon
dioxide) that separate the intermediary metal layers, and finally a passivation layer.
The input electrodes are shown and labeled as nodes V1 and V2. These electrodes
sense a net capacitance through the series combination of the passivation layer
capacitance (CPSV ) and input capacitance (CIN).
The passivation is the unmodified dielectric layer that comes from the CMOS
foundry and generally consists of a stack of silicon dioxide and silicon nitride. Here
we model this layer with as silicon nitride only, with a dielectric constant of 7.5.
Per foundry parameters, the thickness of the passivation is 1.6 ?m, giving an ap-
proximate capacitance area density of 42 aF/?m2. The input capacitance, CIN ,
is the sensed capacitance that couples through the fringing electric fields between
the two input electrodes. As a first-order approximation, this input capacitance is
modeled as a parallel plate capacitance. Given dielectric constants of 50 and 80
for cell cytoplasm and cell media, respectively, the capacitance area density of CIN
is expected to be 92 and 148 aF/?m2, respectively. Since CPSV and CIN are on
the same order of magnitude, they both contribute to the sensed capacitance that
is detected through the CMOS sensor. Indeed, since they are in series with one
another, one way to improve sensitivity to CIN would be to decrease the passivation
thickness or use a high-K dielectric.
An additional double-layer capacitance (CDL) exists at the interface between
the solid substrate and aqueous ionic cell media. However, the size of this capaci-
tance is inversely proportional to the Debeye length which is on the order of several
nanometers for cell media. This means the double-layer capacitance would be orders
22
of magnitudes larger than CPSV and CIN , and so can be ignored [54].
Further parasitics exist at the electrodes that give rise to a baseline offset ca-
pacitance. These include the electrode-substrate area (CP?pp) and fringe (CP?fr)
capacitances, and the metal-metal coupling capacitance between the two electrodes
through the inter-layer dielectric (CMM). Based on foundry parameters, these para-
sitic capacitances are expected to be roughly 8 aF/?m2, 31 aF/?m, and 74 aF/?m,
for CP?pp, CP?fr, and CMM , respectively. The net sensed capacitance between the
two electrodes is given by:
? CP?pp + CP?fr CPSVCINCV 1?V 2 = + CMM + (3.1)
2 CPSV + 2CIN
This two-electrode model can be expanded to consider electrodes laid out in
an interdigitated fashion, as is done on the designed CMOS sensor, and shown in
Fig. 3.2.
CIN
Passivation CPSV
Inter-layer
dielectric CP-fr CMM CP-fr
2CP-pp V1 V2 2CP-pp
Substrate
Figure 3.2: A model of the capacitances at the cell-substrate interface with interdigitated
electrodes.
23
The sensed capacitance for this interdigitated case can be approximated by:
CP?fr 3CPSVCIN
CV 1? ?V 2 = CP?pp + + 3CMM + (3.2)
2 CPSV + 2CIN
As can be seen, there is a roughly 3? increase in the sensitivity to CIN ,
althrough parasitic baseline capacitances are also increased. A plot of the sensed
capacitance vs. relative permittivity is shown in Fig. 3.3a. A linear fit of the data
points results in a slope of 12 aF/r and offset of 10.6 fF at r = 0.
3.2 Multi-physics Simulations
To validate this model, the interdigitated electrode (four fingers) cross-section
was created in an finite element method (FEM) multi-physics solver (COMSOL
Multiphysics). Each finger of the electrodes was 4.2 ?m wide and 23.1 ?m long and
spaced 4.8 ?m apart. The thickness of the silicon, silicon dioxide, silicon nitride,
and metal are 50, 4.06, 1.60, and 1.00 ?m, respectively. The cell media was modeled
as a large rectangular volume directly above the passivation layer with r = 80 and
? = 1.5 S/m. The cell was modeled as a 1 ?m tall cylinder with a radius of 30 ?m
laying on top of the passivation layer above the electrodes (r = 50 and ? = 0.2
S/m). A plot of the effective capacitance across the electrodes as a function of the
permittivity of the cell is shown in Fig. 3.3b. A linear fit of the data points results
in a slope of 19 aF/r and offset of 23.4 fF at r = 0.
As expected there is a larger parasitic capacitance component at the electrodes
in the FEM results when compared to the analytical model. This is due to the
24
a) b)
11.8 25.1
11.6 24.9
11.4 24.7
11.2 24.5
?C ? 370 aF ?C ? 570 aF
11.0 24.3
10.8 24.1
45 55 65 75 85 45 55 65 75 85
Relative Permittivity Relative Permittivity 
Figure 3.3: Sensed capacitance as a function of the relative permittivity of the environment
over the input electrodes, for a) the analytical model, b) the FEM simulation.
simplistic parallel plate modeling of CIN that ignores the fringing effects that will be
the primary contributor in co-planar electrodes. The overall change in capacitance
due to a cell binding event, where r decreases from 80 to 50 is ?C ?= 570 aF and
370 aF for the FEM and analytical models, respectively. Thus the CMOS sensor
must be able to resolve capacitance changes well below this threshold in order to
reliably detect cell binding interactions.
3.3 Capacitive Sensing Approach
A custom CMOS capacitance sensor has been designed based on the capacitance-
to-frequency transduction mechanism. The sensor and overall system was designed
to have a large input dynamic range, provide unsupervised high resolution measure-
ments, require a minimal number of I/O pins, and have a readout interface that
requires only low cost components and allow for multiple devices.
25
Capacitance (fF)
Capacitance (fF)
3.4 Sensor Element
This section describes the individual sensor element that has been developed.
Work related to the sensor was presented in two prior conference publications [55,
56], and journal paper [57].
3.4.1 Oscillator
The basic element in the capacitance-to-frequency transduction mechanism is
the oscillating signal generator. There are numerous oscillator topologies that can
be used, including LC tank, RC, and ring oscillators. These different structures
have varied advantages and disadvantages. The ring oscillator topology was used
for its simplicity, small size, and stability.
The basic sensor cell consists of a voltage-controlled oscillator (VCO) made of
three stages [51, 55, 56, 57] and can be seen in Fig. 3.4a. It has two of its nodes
connected to electrodes fabricated in the top-most metal layer. The capacitance
sensed on these electrodes load the oscillator, causing changes in its charging and
discharging rates, leading to a change in frequency at the output [55]. The parasitic
load capacitances at each inverter stage in the oscillator (CL1?CL3) introduce finite
switching delays (?1 ? ?3) and these contribute to the total delay period which sets
the oscillation frequency of the sensor:
IB
f = (3.3)
VTH(CL1 + CL2 + CL3)
where VTH is the MOSFET threshold voltage, and IB is the bias current. The
26
CIN
VB
V
IB IB I
O
B Controller & I2C
4
EN Pixel Array Pixel SCL
Select
16 32 SDAMUX
Count Data DRDY
16:1
In CLK
RESET
(a) (b)
Figure 3.4: a) Circuit diagram of capacitance sensor pixe, b) System block diagram
interdigitated sensing electrode contributes an input capacitance, CIN , across one
stage of the ring oscillator. During a switching transition, one side of the capacitor
is charged to VDD while the other is discharged, increasing the effective capacitance
across the inverter through the Miller effect. The delay period of this input stage
can be written as [58]:
VDD
?2 = (CL2 + 3CIN) (3.4)
IB
This means a hyperbolic change in frequency is expected as the input capac-
itance changes. However, under the expected operating conditions and expected
small variations in sensed capacitance, the capacitance-frequency relationship can
be approximated as linear, which is discussed in the following sections. On the
fabricated sensor, the NMOS transistors making up the oscillators are sized with a
width-to-length ratio of 2.8/0.7 ?m.
27
Microcontroller
3.4.2 Simulation Performance
3.4.2.1 Capacitance Sensitivity
Fig. 3.5 shows simulation results of how the output oscillator frequency varies
as a function of sensed capacitive load. As expected from analytical equations, the
oscillation frequency decreases as the sensed capacitive load increases. The expected
range of capacitance load values based on an FEM solver model was 2 fF with a
baseline of approximately 10 or 20 fF, depending on the model. Within this range,
the output frequency varies quite linearly, with an R2 value of 0.9985 and a slope of
1.22 MHz/fF.
Ring oscillators exhibit variations in their outputs due to noise sources such
as flicker and white noise [59]. The jitter due to this resulting phase noise must be
considered because it may affect the final output count. Simulations incorporating
flicker and white noise showed an approximately 24 ?s accumulation of jitter over
the 1 s of collection time. Given the sensitivity of 1.22 MHz/fF at an oscillation
frequency of 56.6 MHz, the minimum resolution limited by phase noise comes to 1.1
aF. This figure is an improvement over previous designs, but is based upon expected
sensitivity obtained from simulations.
Monte Carlo simulations of mismatch variations were performed to see their
effects on the oscillation frequency. Fig. 3.5 shows results for a range of sensing ca-
pacitances. The standard deviation in oscillation frequency at a sensed capacitance
of 10 fF is 759 kHz. This leads to an expected uncertainty of 1.8 fF in measurements
due to mismatch. This mismatch is greater than many cell-related capacitances of
28
Figure 3.5: Output frequency as a function of sensed capacitive load. Results for typical
expected capacitances are shown in red. The dotted line shows Monte Carlo simulation
results with the inset showing a histogram for a sensed capacitance of 10 fF.
interest, so the system will require per-pixel calibration so that it can retain sensitiv-
ity to cellular phenomena. Fortunately, in this counter-based design, saturation and
dynamic range are not issues, so the calibration will be implemented in post-readout
digital signal processing.
3.4.2.2 Temperature Sensitivity
A series of temperature sensors were implemented on chip in addition to ca-
pacitance sensor elements. In order to simplify the overall system, the temperature
sensors were designed as additional pixels in the chip using the same base ring os-
cillator. Instead of a fixed reference current, a bandgap reference circuit is used to
bias the ring oscillator with a current that is proportional to absolute temperature
[60].
Fig. 3.6 shows how temperature affects the simulated output frequency of
both the capacitance and temperature sensors. The temperature range of interest
29
Figure 3.6: Output frequency for the temperature and capacitance sensors as a function
of temperature.
is from 36 to 40 ?C, since the experiments will be done in a temperature-controlled
incubator. The output frequency of the temperature sensor responds linearly to
changes in temperature, with an R2 value of 1.00 and a slope of 1.47 MHz/?C. The
capacitance sensor is also affected by ambient temperature, and this relationship
is also linear, with an R2 value of 1.00 and a slope of 160 kHz/?C. These linear
relationships allow for pixel-level calibration before experiments for correcting the
output data.
3.5 System Overview
The lab-on-CMOS chip contains an array of 4?4 sensor elements. In order to
facilitate the required automatic data readout without the use of large and expensive
laboratory equipment, a serial readout bus is implemented. The system is designed
to allow for simple implementation and ease of experimentation. It only requires
a microcontroller and computer for data readout and storage. Fig. 3.4b shows a
30
block diagram of the system. Each of the 4?4 sensor pixels generates an oscillating
signal. This is passed through a digital buffer and into a multiplexer which then
connects to a 32 bit counter. The counter integrates the oscillating signal for a
specified time. Once the integration time is reached, the final count value is stored
and then sent off-chip via an I2C interface implemented on chip. Details of the logic
can be found in Section A.1. In addition to the capacitance sensor pixels in the
sensing array, two additional sensors are also included on chip placed within the
outer padframe area of the chip. This isolates them away from the wet environment
during experimentation and allows them to be used as reference pixels, as described
in Section 3.10.
3.5.1 Multiplexer and Counter
The output signals from each pixel are buffered to a standard digital multi-
plexer, and the selected signal is fed to a single shared counter, which reduces the
total circuit area. The individual pixels are read out sequentially. The minimum
achievable sensor resolution (?fmin) is inversely proportional to the collection time
(Tc):
Qmin
?fmin = (3.5)
Tc
where Qmin is the desired minimum difference in count value. Therefore, one can
trade sensor speed for resolution. For typical operation the sensitivity of the sensor
is 1.22 MHz/fF, or 12.2 kHz/10 aF (Fig. 3.5). Given a target resolution of 10 aF
and an arbitrary desired minimum difference in count value of 100, the required
31
integration time is 8.2 ms. In reality, however, parasitic capacitances [50] reduce
sensitivity [9]. To address this, the integration time Tc in the final system design
is set to be >1 s, providing a safety margin of two orders of magnitude for the
measurement range of interest. This collection time is set using a timing clock that
is generated by frequency dividing the system clock (fclk). Divisions are in factors
of 2 and so the maximum timing clock frequency that satisfies the integration time
requirement is f 22clk/2 = 0.72 Hz, which means Tc = 1.40 s.
The minimum counter size (Nc) required to avoid saturation of the sensor
signal is a function of the oscillation frequency and counting time:
Nc = ceil(log2(fTc) (3.6)
where f is the sensors oscillation frequency. Given a frequency of 56.6 MHz and
collection time of 1 s, the minimum counter size is 26 bits. The I2C serial readout
protocol used in this system has packet sizes of 8 bits, so a 26 bit data transfer
requires 4 packets, where the last byte is zero-padded. Therefore, a 32 bit counter
is used in this design. The added space and power costs due to this increase are
minimal. Additionally, the larger counter, having a maximum Tc of 75 s, allows for
more flexibility in varying the collection time and leads to a minimum theoretical
resolution of 0.06 aF. In reality, phase noise would limit resolution to a higher floor.
32
3.5.2 Serial Readout
Each sensor pixel generates a 32 bit data sample. The most direct method of
reading out the samples is to output each bit to an I/O pad. However this results
in 32 bond pads that must be dedicated for readout, which may not be feasible
in real fabrication. Additionally, a large number of I/O pins introduces packaging
difficulties, especially in cell culture monitoring for which passivation of the electrical
leads is crucial. An alternative method is to read subsections of the 32 bits at a
time. For example, if one byte of data is read out at a time, the number of bond
pads required for readout would be 10 (eight bit lines and two byte select lines). To
minimize the I/O requirements further, we implement an I2C serial readout bus.
I2C is a serial bus data transfer protocol designed for communication between
multiple sensor chips. It implements a two-wire bus with support for up to 128
individual peripheral chips. The I2C communication interface was implemented by
modifying an open-source I2C core [61] written in Verilog HDL. The core code was
modified to implement only read commands with a packet size of 4 bytes. A separate
logic controller was created to interface with the I2C block. This logic controller
implements the sequential reading and data transfer phases of the system. The
oscillator in each pixel is turned on, and the signal is fed into the shared counter.
Once the required counting time has been reached, the value is latched and sent to
the I2C controller for transmission off-chip. The oscillator start-up time may vary
between pixels, so a delay stage has been implemented to allow the oscillator to
start up before the counter starts accumulating. Another consideration is the finite
33
rise and fall times of switching bits in the counter, which introduce brief (< 100
ps) periods of time when the count is unstable and should not be latched. In this
implementation, latching only occurs after a negative edge of the oscillator clock,
since the counter is positive-edge triggered. The digital I2C and chip controller logic
were synthesized using an RTL compiler. A floorplan layout was then created using
Place & Route software and merged with the manually drawn pixel array.
3.6 Fabricated Chip & Readout Platform
The 4? 4 array of pixels (shown in Fig. 3.7) were arranged with an X and Y
pitch of 196 ?m and 186 ?m, respectively, covering an area of 618 ? 588 ?m2 on
a 3 ? 3 mm2 chip. Since the electrodes are 30 ? 30 ?m2 in size and are sensitive
to permittivity changes in their immediate surroundings, the actual sensing area
covered by the electrodes is roughly 120 ? 120 ?m2. This means that the spatial
resolution of the sensor is relatively low. However, when looking at the mechanics of
the cell population as a whole, discrete measurement sites spread over a large area
can suffice. A higher pixel density would be ideal for performing studies of individual
cells. Eight I/O pads are used which were duplicated for packaging redundancy.
The readout system used in this work is a commercially available microcon-
troller (MicroPython PYB v1.0), although any device capable of I2C communication
can be used. All bias voltages are generated internally (Section A.1) and the supply
voltage of 3.3 V is provided by the microcontroller directly. The finite ramping time
of the power supply and startup circuits built into the bias generator ensure the
34
oscillators start up correctly. The microcontroller additionally supplies the digital
clock signal that is used for timing of the controller logic on chip and the I2C bus.
Therefore no additional equipment besides the sensor, microcontroller, and a battery
are required to perform measurements, allowing for experiments to be done easily
outside of the laboratory setting. The average power consumption for the entire
system was measured to be 300 mW with the IC chip consuming 8 mW. Therefore
the complete lab-on-CMOS system can easily be implemented on a portable plat-
form. In fact, a low power microcontroller can be used in order to extend battery
lifetime to several days to facilitate longer term experiments.
The presented sensor was fabricated in a commercially available 0.35 ?m
CMOS process and wirebonded to a standard DIP-40 ceramic carrier for dry bench-
top experiments. For wet experiments, the wirebonds were passivated using a high
electrical resistivity epoxy (Durapot 863, Cotronics). The passivation was performed
by first heating the chip-in-carrier to 120 ?C and then manually applying the epoxy
(a) (b) (c)
Figure 3.7: Photomicrograph of a) Fabricated die (3 ? 3 mm2), b) sensing array, and c)
sensing electrodes (30 ? 30 ?m2).
35
Metal Probe
IC Chip
Chip Carrier
(a) (b)
Figure 3.8: a) Exeprimental setup to characterize sensor as a distance sensor. b) Sensor
output frequency as a metal probe is moved away from the surface of the chip. Solid line
shows a regressive fit of the measured data.
over the bond wires and onto the chip surface, leaving only the sensing area uncov-
ered.
3.7 Capacitance Computation
The sensor data was calibrated by placing a metal probe close to the surface
of the chip and varying the vertical distance of the probe relative to the chip (Fig.
3.8a). The position of the probe was controlled automatically using a piezoelec-
tric actuation platform. The probe was moved in steps of 5 ? 0.5 ?m, and the
sensor outputs were allowed to settle for 10 minutes between movements. These
experiments were performed with Sheung Lu. Fig. 3.8b shows how the measured
frequency of oscillation varies as the metal probe is moved away from the chip sur-
face. The output frequency increases as a function of probe distance, corresponding
to a decrease in capacitance.
36
Probe Distance 
(?m)
This setup was simulated in a FEM solver to obtain the capacitive loads in-
duced by the probe positions. For the small operating range of interest (less than 2
fF) in cell sensing applications, the sensor output frequency response, f(CIN), can
be modeled linearly as follows:
f(CIN) = ??(CIN + C0) + f0 (3.7)
where CIN is input capacitance, C0 is the total parasitic and offset capacitances at
the circuit level, f0 is a baseline frequency, and ? is the sensor?s sensitivity. The
capacitive load measured due to the metal probe is then modeled using an inversely
proportional relationship:
?
CIN(x) = ? (3.8)x+ x
where ? is a scaling parameter that relates sensed capacitance to probe distance,
and x? is a parameter that accounts for any displacement errors introduced by the
piezoelectric actuators. We then obtain a model of the sensor?s output frequency as
a function of probe distance:
? ?f(x) = ?( ? + C0) + f0 (3.9)x+ x
The parasitic capacitances in this model were estimated to be 12.1 fF by simu-
lating the electrode and probe structure in an FEM solver (COMSOL Multiphysics).
A regressive fit was performed on the data and the resulting curve is shown in Fig.
3.8 (solid line). The sensor output can thus be characterized by generating calibra-
37
59.6
59.4
59.2
59.0
59.6
59.4
59.2
59.0
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
Capacitance (fF)
Figure 3.9: Sensor characterization curves generated by mapping electrode probe distance
to modeled capacitance value. The curves from left to right indicate the best, average,
and worst case sensitivity scenarios.
tion curves using this relationship. Fig. 3.9 shows three curves corresponding to
the best, average, and worst sensitivity. The capacitance sensitivities corresponding
to each scenario are 1077, 593, and 171 kHz/fF. respectively. These curves were
obtained by using the smallest, largest, and average ? values obtained from a set of
1000 regressive fits. The sensitivity figure ? = 593 kHz/fF, is the measure that is
used in subsequent capacitance calculations.
38
Frequency (MHz)
3.8 Bench-Top Characterization
3.8.1 Response to Organic Solvents
The packaged chip was evaluated by measuring the response to fluids placed
in the sample well. We selected fluids with a range of dielectric constants in order
to induce changes in permittivity of the environment surrounding the chip, causing
changes in signal output. The fluids were deionized (DI) water (r = 80.1), acetone
(r = 20.7), and isopropyl alcohol (IPA) (r = 17.9) [62].
Cell media was also used. Fig. 3.10a shows data from a representative pixel
that illustrates the application and removal of each fluid over time. Fluids with
higher relative permittivity induce a stronger drop in amplitude from the baseline
(air, r = 1.0) since they correspond to a larger sensed capacitance. Of the tested
solutions, DI water has the highest relative permittivity and causes a change in
capacitance of 12 fF from the baseline, based on FEM simulations. Therefore the
input range of the sensor as tested in this work is 12 fF. The sharp features seen
in Fig. 3.10a are artifacts related to physical movements in the experiment. This
is especially the case with acetone since its relatively high vapor pressure means
constant fluid replenishment is needed if the package is not sealed. Fig. 3.10b
shows the mean sensor output across all 16 pixels with error bars marking standard
deviations, and Fig. 3.10c shows the response as a function of relative dielectric
permittivity.
The capacitance sensor chip was also exposed to varying concentrations of
IPA diluted in DI water. The bottom panel of Fig. 3.11 shows the sensor out-
39
a) Cell media
100
IPA Methanol Acetone
80 Water
60
100 150 200
Time (min.)
b)
75
70
65
60
55
IPA Methanol Acetone Water Cell Media
c)
75
70
65
60
0 20 40 60 80 100
Relative dielectric permittivity
Figure 3.10: Sensor response to addition of different fluids. (a) Transient plot of response
for a single pixel. (b) Mean response and standard deviation across 16 pixels. (c) Mean
response as a function of relative dielectric permittivity.
put frequency, averaged across 15 pixels; data from one channel was discarded due
to readout error. As can be seen the sensor frequency response decreases as the
percentage by volume of IPA is decreased.
40
Frequency (MHz)
3.8.2 Spatial Correlation of Capacitance Sensors
The sensor chip contains an array of 16 capacitance pixels, distributed into four
rows and four columns. Following the experimental protocol outlined in Section 3.7,
it is possible to create a map of responses over time by matching each pixel with
its physical location. Fig. 3.12 shows an experiment where the metal probe was
brought down onto the sensor surface and then lifted off. For this particular run,
the probe was purposefully aligned off-center from the sensor array. The change
in frequency (?f) from the baseline measurement is plotted for all 16 pixels. The
bottom panel shows a map of the 16 responses at four different time points, marked
by the blue dashed line in the top panel. Each grid square corresponds to a pixel?s
physical location on the chip. The probe begins to touch down at 190 min and
lifts off at 290 min. The bottom panel shows that the change in frequency is more
prominent in the right side and upper-right side pixels. As the probe lifts off, the
Figure 3.11: Sensor response due to application of difference concentrations of isopropyl
alcohol.
41
(a)
(b)
Figure 3.12: Mapping sensor response to external probe interacting with chip surface. a)
16 pixel responses plotted over time as the probe is brought down and lifted off. b) Shaded
plot mapping responses to physical pixel locations. Each array image corresponds to time
points indicated by the dashed lines in (a).
pixels gradually decrease in their response, with the top-right pixel being the last
to settle. This is consistent with the initial alignment of the metal probe in this
experiment.
42
3.9 System Level Improvements
3.9.1 Parallel Readout
In the 2nd generation of the chip, the architecture was improved to perform
parallel readout. The chip maintained a 4? 4 array, with multiple sensors activated
at the same time, in a column-wise format. Four row units, each consisting of a
multiplexer and counter, were used to integrate a column of sensors at one time.
Each communication cycle on the I2C readout bus would then send 4 ? 32 bits of
data. Fig. 3.13a shows a block diagram of the system with a column-wise readout.
This version allows for a 4? improvement in sensor speed allows the IC to be used for
shorter-term cell experiments while maintaining the same high resolution required
for detection of cell binding events. A die photo of the fabricated chip is shown in
Fig. 3.13b.
RU
Controller & I2C
RU
Activate
SCL
RU
SDA
RU
DRDY
CLK
MUX
Count
4:1 RESETData In
Row Unit (RU)
(a) (b)
Figure 3.13: a) Expanded system to include column-wise readout. b) Die photo of the
2nd generation chip
43
Microcontroller
3.9.2 Tunable Integration Time & Active Pixel Selections
The resolution of the sensor is directly related to the integration time, as
discussed in Section 3.5.1. In the generation 1 fabricated chip, the integration time
was set to 1.39 seconds (1/(f 22clk/2 ), fclk = 3 MHz) per pixel, which means 22.2
seconds are required to read 16 pixels. This is a relatively slow sampling rate, but
while it is adequate for cell experiments with long time constants, a faster sampling
rate would help to capture minute differences that would otherwise be lost. The
most direct way to increase sampling rate is to reduce the integration time at the
cost of reduced resolution.
To this end we look at the noise resolution of the device. The sensor was
run under stable operating conditions (sensor surface covered, dark light conditions)
over a period of several hours. The mean standard deviation of the output frequency
across all channels was computed to be 22.3 kHz which means a noise floor of 37.6
aF (given the measured sensitivity of 593 kH/fF). For an integration time of 1.39
s, the 22.3 kHz noise level means that a value of ?62?103 of the integrated sensor
count output can be considered within the noise floor, that is the effective number
of bits (ENOB) is 16.1. Thus, by reducing the integration time, the ENOB can be
increased. At the extreme, an integration time of ?22.4 ?s would reduce the noise
value to ?1.
With this in mind, a 3rd generation chip was designed and fabricated. The
chip is designed with a programmable integration time from 20 ms to 5.6 s, in 20
ms increments. This parameter is programmed through the I2C bus and does not
44
require any additional wiring requirements. This version of the chip also implements
programmable sensors to activate and deactivate sensors in the array individually.
This spatial readout selection allows the sensing area of the array to be fine-tuned
for certain applications. For example, if global capacitance variations are expected
such as during cell death and detachment due to application of a toxin, the number
of active pixels can be minimized in order to increase sensor sampling rate. An
increase in sampling rate of 32x is achievable if only a single sensor is required for
readout. A block diagram and die photo of this 3rd generation chip is shown in Fig.
3.14.
Finally, another aim is to enable high throughput bioassay experiments by
allowing multiple devices to run in parallel without the need for additional equip-
ment. This could prove to be useful in certain applications, and help to make this
lab-on-CMOS system a more practical tool for biological assays.
Logic 
Bias Gen.
Controller
Pixel Array CLK
2
Count I C Serial 
RST
Interface
DRDY
Row/Column Integration SCL
SRAM
Decoders Time Coeff.
SDA
I2C Address
(a) (b)
Figure 3.14: a) System block diagram of chip with programmable integration time, pixel
activation and programmable I2C address. b) Die photo of the 3rd generation chip
45
MUX
3.10 Sensor Calibration
Pixel-to-pixel variations in the output frequency are expected due to process
variations and parasitics inherited from fabrication. This can result in each of the
sensor pixels having slightly different baseline frequency levels from one another.
Additionally, fluctuations in power supply levels and temperature may affect the
pixel output frequencies. Therefore, measurements are computed with respect to
an on-chip reference pixel that is also exposed to global supply and temperature
variations, but is unaffected by the analytes being detected. This provides a pseudo-
differential measurement as given by the following:
?CN(t) = ?([fN(t)? fN(0)]? ?[fRef (t)? fRef (0)]) (3.10)
where fN(t) is the raw data for channel N , fRef (t) is the data for the reference
channel, and ?CN(t) is the final computed capacitance value. The parameter is the
frequency-to-capacitance sensitivity factor (590 kHz/fF) obtained experimentally in
Section 3.7, [56]. The parameter ? is a calibration coefficient that can either be
fixed for all channels (? = 1), performing a global calibration, or be set individually
for each pixel (? = ?N/?Ref ). Coefficients ?N and ?Ref are the slopes of linear fit
lines for fN(t) and fRef (t), respectively.
3.10.1 Temperature Sensitivity
The temperature sensitivity of the device was analyzed by pipetting a drop
of cell media onto the sensor array and placing the packaged device in a hot plate
46
oven at 44 ?C. Measurements were started and the oven was turned off to allow it
to cool down to 33 ?C over a period of 80 minutes. The mean output frequency of
the pixels is shown in the top panel of Fig. 3.15 (black circles). The temperature-
response of the device is approximately linear (R2 = 0.99) with a sensitivity of 80.6
kHz/?C or 1374 ppm/?C. After calibration using the reference pixels as described
previously with ? = 1, the sensitivity is reduced to 10.1 kHz/?C or 172 ppm/?C.
The resulting data points are shown as red diamond markers. Per-pixel calibration
was also performed by computing the ?N coefficients for each channel. This results
in diminishing temperature sensitivity down to mean across channels of 0.0 kHz/?C.
The data points for this case are shown as the yellow square markers. It should be
noted that in cell culture experiments, the growth environment has to be strictly
controlled and so the ambient sensor temperature is expected to remain stable at
37 ?C throughout the duration of an experiment.
3.10.2 Power Supply Sensitivity
Portable biosensor applications can be prone to power supply fluctuations and
so we analyze the power supply sensitivity of the device. The chip was powered
using an external DC power supply (SourceMeter 2400, Keithley) and the supply
was swept from 3.0 to 3.3 V in 10 mV increments. The mean output frequency of
the pixels is shown in the bottom panel of Fig. 3.15 (black circles). The response
of the device is approximately linear (R2 = 0.97) with a sensitivity of 797 kHz/V.
After global calibration, the sensitivity is reduced to 58 kHz/V (red square markers).
47
Figure 3.15: Sensitivity and compensation for temperature (top) and power supply (bot-
tom)
Finally, the single pixel calibration was seen to reduce sensitivity down to 0.0 kHz/V
(yellow square markers).
48
3.10.3 Noise
Ring oscillators are subject to flicker, thermal, and white noise that result in
phase noise and timing jitter [59, 63]. These effects are manifested as fluctuations in
the oscillators? output frequencies. In order to quantify noise and resolution of the
sensor elements, we performed benchtop experiments under cell culture conditions.
A standard microbiology tube (1.5 mL, Eppendorf Flex-Tube) was cut and
glued onto the packaged chip to form a well. Cell culture media was put into the
well and the device was placed in an incubator at 37 ?C for 15 hours while sensor
data was being recorded. The pixels oscillate with baseline frequencies ranging
between 58.0 and 60.5 MHz. Global calibration was performed (? = 1) as described
in Section 3.10. A plot of the resulting data is shown in Fig. 3.16a. The gray traces
are signals for each pixel while the black line shows their mean. Fig. 3.16b. shows a
histogram of the mean output. The mean standard deviation of frequency over all
pixels is 9.3 kHz, which corresponds to a dynamic capacitance variation of 16 aF.
It should be noted that while pixel level calibration reduces temperature and
power supply sensitivity, it requires computation of calibration coefficients. There-
fore for simplicity, the in vitro experimental data presented in this work uses global
calibration (? = 1). As such, sensor measurements are expected to exhibit noise
contributions from random temperature and power supply fluctuations. However,
the extent of the noise contributions is expected to fall within the measured noise
resolution of 16 aF. For example, since experiments are done in a regulated incuba-
tor, temperature stability is expected to be within ?0.2 ?C [64] which corresponds
49
to ?3.4 aF given a 17 aF/?C sensitivity. Additionally, the chip is powered from a
voltage regulator on the microcontroller board, with a dropout voltage of < 50 mV
at a 2.5 mA load current [65]. This corresponds to variability of roughly ?5 aF
given a 98 aF/V sensitivity.
 
 
 
a) 
 
b) 
 
Figure 3.16: Long-term sensor response and noise when exposed to cell media at 37 ?C.
a) Output frequency vs. time for each pixel (gray) and their mean (black). b) Histogram
of the mean frequency.
50
Chapter 4
Tracking Cell Metrics
4.1 Introduction
This chapter presents results of in vitro experiments testing the capacitance
sensor chip with live cells. The experimental protocol is described, along with discus-
sion of encapsulating the chip in biocompatible materials. Experimental results are
then presented with preliminary ground-truth validation using standard microscopy.
Finally the sensor data is used to corroborate with actual cell mechanics by gener-
ating quantification metrics to monitor cell adhesion, proliferation, and mitotis.
4.2 Device Packaging
The presented sensor was fabricated in a commercially available 0.35 ?m
CMOS process Three different packaging methods were used. The first method
involved gluing the dies to a ceramic carrier, wirebonding to form electrical con-
nections, and passivating using an epoxy, as described in Section 3.6. The second
method involved encapsulating the die in an epoxy mold, patterning a zinc/copper/nickel
metal stack to make electrical contacts with the chip, and finally passivating the
electrical traces using a second layer of epoxy [66]. Cross-sectional schematics of
these two packaging methods are shown in 4.1. The third method involved mount-
ing the chip to a low-temperature co-fired ceramic (LTCC) carrier by using Au-Au
51
a)
passivation nickel
copper
chip
epoxy
b) passivation wirebonds
chip
DIP-40 ceramic carrier
Figure 4.1: Cross-sectional views of chips packaged in a) an epoxy handle wafer, and b) a
passivated DIP carrier
thermocompression bonding. Under-filling with epoxy was used to create a liquid
tight seal between the sensor chip and the LTCC carrier [67, 68]. For all packaging
methods, cell culture wells were created by gluing either standard microbiology or
polypropylene tubes around the sensing area. No variation in chip characteristics
or performance were observed due to choice of packaging method. Photographs of
chips in each of the three packages are shown in Fig. 4.2a (from left to right: ceramic
carrier, epoxy carrier, LTCC carrier).
The readout system consisted of two printed circuit boards (PCBs), a daugh-
terboard that housed the chip and a motherboard that held the microcontroller
(MicroPython PYB v1.0) used for readout (left panel of Fig. 4.2b). The daugh-
terboard contained a zero-insertion-force socket that the chip plugged into. The
52
motherboard contained headers that interfaced with the microcontroller. A flat
shielded Ethernet cable was used to connect the two PCBs. The microcontroller
stored the data locally on an SD card and also sent the data to a graphical user
interface on a laptop computer (custom MATLAB GUI) which plotted and stored
the data in real time. Power to the chip was supplied by a voltage regulator on the
microcontroller, which itself was powered via USB from the computer. A photo-
graph of the microcontroller, and daughterboard inside the incubator is shown in
the right panel of Fig. 4.2b.
4.3 Experimental Protocol
The results shown in this chapter are from experiments performed with fab-
ricated sensors. These peliminary biological tests were performed by growing ad-
herent cells onto the chip array. Growth media (RPMI 1640) was prepared with
10 % fetal bovine serum, and supplemented with antibiotics and anti-fungals (peni-
cillin/streptomycin 100?). Human ovarian cancer cell lines A2780 and CP70 were
used in this experiment. The cells were grown in a cell culture vessel until they reach
lag phase growth. At this point the cells were detached using 0.25 % Trypsin/EDTA,
formed into a pellet and re-suspended into 12 mL of fresh cell media. The devices
were prepared by sterilizing them with UV light, and rinsing them with deionized
water, phosphate buffer solution (Dulbeccos Phosphate Buffer Solution), and cell
media. Fresh media was then added and the device was maintained in an incubator
(37 ?C, 5 % CO2) for 15 minutes. Then 40 ?L of the cell suspension was added to
53
(a)
(b)
Figure 4.2: a) Photographs of the capacitance sensor chip packaged using the three dif-
ferent methods (from left to right: ceramic carrier, epoxy carrier, LTCC carrier). b) The
test setup. Left: image of a packaged chip on a daughterboard connected to the readout
microcontroller using an Ethernet cable. Right: image of the daughterboard inside the
incubator.
54
the device, and data was continuously recorded throughout the experiment. Two
separate devices were packaged and plated with each of the two cell lines in separate
experiments, resulting in a total of four experimental datasets. All 16 sensor pixels
on each chip were active and recording. The devices were kept in the incubator and
connected to the readout microcontroller and laptop (located externally) using a
shielded Ethernet cable. Fig. 4.2b shows a photograph of the testing setup.
4.4 Tracking Cell Adhesion & Migration
Fig. 4.3a shows response curves obtained from a device plated with CP70
cells. The two vertical lines at t = 21 hours and t = 44 hours indicate time points
when recording was paused to take images of the chips surface, shown in Fig. 4.3b
and Fig. 4.3c. For all but one sensor the signal remained relatively low until 5 hours
of incubation. This is consistent with the pre-adhesion phase of growth where cells
settle onto the chip surface. After this phase, cells start adhering to the chip surface,
showing a corresponding increase in signal. The strength of the signal change is
indicative of cell growth quality at that electrode, which in turn is dependent upon
where the cells first settle down from the solution. Therefore certain electrodes show
little change in signal if no initial settling (and subsequent adhesion) occurs. This is
evident in sensor 5 (row 2, column 1 of the array) which shows little change through
34 hours of incubation. Fig. 4.3b shows the corresponding absence of nearby cells
at this electrode. Likewise, the response of sensor 11 (row 3, column 3) stays below
50 aF through 44 hours of incubation. Both Fig. 4.3b and Fig. 4.3c show this
55
electrode to have no cells directly on top of the sensing electrodes.
Fig. 4.4 shows results from a second and third experiment, using the same
device. Each panel shows the output of two sensors, one that was observed to have
cell coverage (red) and one without coverage (black). As before, the results show
a correlation between sensor output and observed cell coverage. A difference of at
least 500 aF was observed between electrodes with and without cell coverage.
As cells adhere and proliferate, they can also translate from one location to
another, a phenomenon called cell migration [69]. This can also be observed in the
transient response curves of Fig. 4.3a. As cells proliferate and cover more electrode
area, a corresponding increase in absolute sensed capacitance is observed. Likewise if
cells migrate away from an electrode, they cover less and less electrode area, causing
a corresponding decrease in signal. Several sensor channels show signal fluctuations
600 1
400 2
200 3
4
0
600 5
400 6
200 7
8
0
600 9
400 10
200 11
12
0
600 13
400 14
200 15
16
0
0 5 10 15 20 25 30 35 40 45
Time (hours)
(a) (b) top (c) bottom
Figure 4.3: a) Response curves of the chip as CP70 cells adhere and proliferate across the
sensor array. The two vertical black lines at t = 21 hours and t = 44 hours indicate times
when data recording was paused for imaging. Right panel shows image of surface of chip
after b) 21 hours, and c) 44 hours of incubation of CP70 cells.
56
|?C| (aF)
Exp. A
Exp. B
1000
Exp. A
750 Sensor with cells
Sensor without cells
500
250
0
1000
Exp. B
750
500
250
0
0 10 20 30 40
Time (hours)
Figure 4.4: Results from two cell experiments performed after cleaning and reusing a
device. Lines show output from sensors with (red) and without (black) cell coverage
up and down over a period of several hours. These trends can be caused by cell
migration to and away from the electrodes. For example, sensor 4 (row 1, column
4 of the array) shows a gradual signal drop from 300 aF to 20 aF between t =
21 hours and t = 44 hours. Examination of the images in Fig. 4.3b,c shows a
corresponding absence of cells on sensor 4 between these two time points. Sensors
57
| C| (aF)
10 (row 3, column 2), and 16 (row 4, column 4) also corroborate this observation,
with decreases of 392 aF and 226 aF, respectively. Additional sample data can be
found in Appendix A.2.
4.5 Quantifying Cell Growth Rates
The growth rate of cells can be monitored by looking at the cumulative num-
ber of sensors (NCS) in the array that record a significant change in capacitance,
designated by a threshold value, during the experiment. Fig. 4.5 shows such a plot
with the capacitance threshold set at 200 aF (> 5 times the noise threshold of 22.3
kHz, Section 3.9.2), for each of the four experiments. For clarity, the increments
were recorded for each hour of incubation. As can be seen, there is an initial 5 hour
period where a small number (< 5) of sensors show a capacitance change. This is
indicative of the cell sedimentation process. Then as the cells adhere and proliferate,
the number of sensors showing a response greater than 200 aF increases. Over the
course of several hours the number of affected sensors increases and plateaus.
A quantitative estimate of the cell growth rate can be obtained by calculating
the rate at which sensors detect cells over the course of the adhesion. The slope of
NCS for each of the experiments in Fig. 4.5 was calculated; the solid lines show the
calculated slope line. A nominal area growth rate can then be estimated by scaling
this growth rate by the known 30 ? 30 ?m2 area of each pair of electrodes. Table
4.1 summarizes the estimated growth rate for each of the experiments.
58
Device 1, CP70 Device 2, CP70
15 a) 15 b)
10 10
5 5
0 0
0 10 20 0 10 20
Device 1, A2780 Device 2, A2780
15 c) 15 d)
10 10
5 5
0 0
0 10 20 0 10 20
Time (hours)
Figure 4.5: Cumulative number of sensors that show an increase in capacitance as a
function of time, for each of the four experiments described in Section 4.3. The large
plateau after 22 hours in (b) is due to solution evaporation.
Cell Line Device Adhesion Time Growth Rate Growth Rate
(hrs) (sensors/hr) (?m2/hr)
CP70 1 11.7 0.43 380
CP70 2 9.8 0.74 670
A2780 1 12.7 0.67 600
A2780 2 7.8 0.87 780
Table 4.1: Cell Growth Rate Estimation
The experimental data shows a clear pattern of cell coverage detection; as
a larger area of a pixel?s electrodes are covered, a corresponding increase in the
sensor signal is recorded. Fig. 4.6b shows the average capacitance change across all
experiments. A linear fit (R2 = 0.94) estimates a response of 100 aF/cell. Error
bars corresponding to the standard deviation of the measurements indicate that
variability is relatively large compared to the mean: ?200 aF for 2 cell coverage.
59
Cumulative number of sensors detecting cells
However, since the experiments vary in both device used and type of cells cultured,
discrepancies are expected.
The dynamic change in measured capacitance also has strong correlation with
changes in cell coverage. Fig. 4.7 shows such a plot for the experiment with Device
1 and CP70 cell line. The change in measured capacitance (?C) is the difference in
signal between two time points that correspond to the two images. The change in
cell coverage is the difference in the estimated number of cells covering each pixel
at the two time points. For this experiment the change in coverage ranged from ?3
to +2 cells.
As can be seen there is a correlation between the change in cell coverage and
a)
Device 1, CP70
800
Device 2, CP70
600 Device 1, A2780
400 Device 2, A2780
200
0
b)
800
600
400
200
0
0 1 2 3 4 5
Estimated Coverage (# cells)
Figure 4.6: Plot of the measured sensor response as a function of estimated cell coverage
of pixels. a) All experiments separately, b) Mean value. Linear fit shows a sensitivity of
100 aF/cell.
60
| C| (aF)
200
150
100
50
0
-50
-100
-150
-200
-250
-300
-3 -2 -1 0 +1 +2
(Cell Coverage)
Figure 4.7: Plot of the measured change in capacitance as a function of change in cell
coverage (Device 1, CP70 cells)
the change in measured capacitance. An average decrease of 300 aF in the signal
was observed for sensors that showed a 3 cell decrease in coverage between the two
time point, and a 130 aF decrease for a sensors exhibiting a single cell decrease in
coverage. However an average signal increase of 140 aF was measured for sensors
that had no change in coverage. We note that the cell coverage estimates were
obtained manually and reflect an estimate of the number of cells on the electrode,
so reported changes in cell coverage are limited in accuracy.
4.6 Summary
This chapter presented preliminary results of in vitro cell experiments using
the cell capacitance sensor. The chip was packaged using three different meth-
61
C (aF)
ods (including one assembled at the University of Oulu) and does not require any
surface functionalization or post-processing prior to experiments. Human ovarian
cancer cell lines CP70 and A2780 were cultured in wells attached to the chip surface,
and capacitance data was taken automatically over a period of three days. Micro-
scopic images were taken of the sensor surfaces during the experiments to provide
visualization of cell growth and proliferation.
Analysis of the data showed transient capacitive changes during experiments
which correlated with visual images of the chip. Estimates of the cell growth rate
were obtained from analysis of the transient data for each experiment. Cell migration
was also observed at some sensor locations. Finally, the sensitivity of the sensors was
quantified by comparing the measured capacitance responses to the cell coverage at
each individual pixel. Analysis of four experiments showed good correlation between
cell coverage and change in capacitance measurement.
In the following chapter, higher resolution imagery is used to further validate
measurements from the chip, including the investigation of single cell binding events.
62
Chapter 5
Real-time Imaging in Lab-on-CMOS Applications
5.1 Preface
This chapter includes excerpts of an article published in IEEE Transactions
on Biomedical Circuits and Systems that was authored by myself, Sheung Lu, Dr.
Marc Dandin, Dr. Elisabeth Smela, and Dr. Pamela Abshire. The core idea for
the two well cell culture platform was formulated in discussions with Sheung Lu.
Sheung additionally sourced parts for the custom microscope assembly. My other
contributions include setting up and performing the experiments, developing the cell
detection algorithm, performing image processing, and most of the writing.
5.2 Introduction
In order to obtain accurate ground truth information, it is important to be
able to obtain visual confirmation of what is occurring on the sensor surface. This is
especially important since capacitive changes are the result of changes in the permit-
tivity of the immediate surroundings at the input electrode, whether it be due to the
adhesion of cells or changes in the media. The quantification of data presented in
Section 4.4 used ground truth information from manual discrimination of cells based
on microscopic images. There are a few drawbacks with this approach. Manual hu-
man identification is inherently subjective since it is difficult to make estimates that
63
are always consistent throughout multiple trials. Therefore a method of objectively
quantifying ground truth cell coverage would be beneficial. A major obstacle that
exacerbates the issue of subjectivity is the difficulty in obtaining clear images of the
sensor surface. Since the sensor substrate is opaque, imaging techniques that require
a through-sample optical path cannot be used. These include phase contrast and
inverted light microscopy that are normally used in microbiology. This means that
reflective microscopy is required. However, acquiring clear images with this method
is difficult due to light refraction through the air-liquid interfaces, as can be seen
in Fig. 4.3b,c. In order to mitigate this problem, higher-quality images could be
obtained using a microscope with immersion lenses [6].
Another important consideration in validating the biosensor is having higher
temporal resolution of the ground truth. This means continual imaging of the sensor
surface as the experiments are run. This introduces logistical difficulties since the
sensor would need to be moved to an imaging facility numerous times through a
recording session. The physical movement itself could compromise the viability of
cells since the samples need to be removed from the cell culture incubator often.
This chapter presents work towards building a platform to perform long-term and
concurrent time-lapse imaging and data collection. No such platform has previously
been reported for LoCMOS sensors.
64
5.2.1 Visualizing Cells
Cells absorb little to none of the light that passes through their bodies. Rather
cells are phase objects: they create changes in image contrast due to variations in
their thickness and their different refractive index from the cell media [70]. As such,
a number of different microscopy methods have been developed to properly visualize
cells. Fluorescence microscopy involves the addition of a fluorescent stain that can
label certain proteins or chemical compounds within cells. These labels are excited
with light at a certain wavelength, and the resulting fluorescence emission can be
measured. However, the dyes themselves can affect normal cellular function and so
are rarely used in long-term live cell imaging [15]. Additionally, the excitation light
intensity can be phototoxic to cells, and affect their behavior over time [70, 15].
Phase-contrast and differential interference contrast microscopy are popular label-
free alternatives to fluorescence imaging [71]. They use additional optical hardware
to convert phase shifts caused by cells into detectable light intensity variations. Most
cell microscopy methods are commonly used with optically transparent samples, and
so are not suitable for imaging cells on optically opaque CMOS chips.
5.2.2 Challenges
There are several significant challenges in obtaining clear images of a CMOS
sensor in the cell culture environment. Since the sensor substrate is opaque, imag-
ing techniques that require a through-sample optical path cannot be used. These
include phase contrast and inverted light microscopy that are commonly used in mi-
65
a) b) 0 hrs
5 hrs
Figure 5.1: Photomicrographs of cells on CMOS chips. Left: Images of sensory neurons
taken with a) standard objective lens, b) immersion lens. Image from [6]. Right: Images
of ovarian cancer cells, with media evaporation resulting in loss of focus after 5 hours.
crobiology. This means that reflective microscopy is required. However, acquiring
clear images for samples in a fluid environment with this method is difficult. Light
refraction through the air-liquid interface results in blurry or poor quality images,
as can be seen in Fig. 5.1a (left panel). Curvature of the interface due to surface
tension (i.e. meniscus) is particularly problematic. Higher-quality images can be
obtained using immersion lenses [6], which are designed to be used with an immer-
sion liquid (e.g. oil). Water immersion lenses, in particular, can be used for live
cell imaging by dipping the lens directly in cell medium, resulting in much clearer
images Fig. 5.1b (left panel). However these lenses are quite expensive and are not
typically used for long-term observation [72].
Mammalian cells require a stable growth environment that matches physio-
logical conditions, usually achieved by keeping the sample inside an incubator (37
?C, 5% CO2). Since a continuous stream of images is required for validating the
66
high-temporal resolution sensor data, the sensor/culture would need to be moved
to an imaging facility numerous times throughout a recording session. This would
compromise the viability of cells since the device needs to be removed from the
controlled culture environment. This often means that the sensor recording needs
to be stopped and restarted as well. Therefore, imaging should occur continuously
within the cell culture environment.
Furthermore, the cell media has to be maintained at a constant pH level for
optimal growth conditions. This is ensured by gaseous exchange with 5% CO2
within the incubator, so the culture well cannot be covered with an airtight seal.
During long-term experiments, this leads to evaporation of the cell media which
increases the salt concentration and osmotic pressure and decreases viability [73].
Media evaporation further causes a change in the height of the air-liquid interface,
leading to the sensor surface moving out of focus, as can be seen in the right panel
of Fig. 5.1. This is not an issue in traditional transmittance-based microscopy,
because the lens is focused on the underside of culture well that always remains the
same distance away from the lens.
5.2.3 Alternative Imaging Systems
Several research and commercial grade systems exist that integrate cell cul-
ture incubator conditions and imaging [74, 75, 76, 77, 78]. These systems generally
incorporate an inverted microscope, illumination sources, and CCD camera into
an incubation chamber with standard heat and humidity controls. The stage po-
67
Borescope 
LEDs 
Imager 
Sensor 
Figure 5.2: Schematic of a borescope-based imaging system. Placement of the lens within
the cell solution can help to reduce effects of light refraction.
sition can either be manually or automatically controlled. Each of these systems
are designed for use with transparent samples since the inverted microscopes rely
on light transmission through the samples. Since substrates on lab-on-CMOS chips
are opaque these systems would not be applicable without modifications. Jin et al.
developed a wireless imaging platform that can be placed inside commercial incu-
bators [74]. It is based on a Sony QX10 camera that his housed in a 3D printed
bracing structure that includes an excitation filter, emission filter and lens for flu-
orescent microscopy applications. The camera itself is WiFi enabled and so it can
be controlled remotely, eliminating the need for wiring that needs to pass out of the
incubator door, which can compromise the stability of the incubation environment.
A low-cost alternative could take inspiration from healthcare. Endoscopes
68
and borescopes are visual inspection devices with a camera head on one end and
an eyepiece on the other. Video versions of these devices act as miniature video
cameras with the eyepiece replaced by either a video screen or connection to a
computer. Additionally, endoscopes in particular are designed to be waterproof. If
the camera-end of the scope can be placed into the culture well, then the problems
of refraction may be avoided. Electronic cameras can be done remotely to allow for
automatic time-lapse imaging (Fig. 5.2).
Another system would involve the use of a normal reflective microscope, with
an attached CCD camera. This would be similar to the setup developed by Jin et al.
[74], with a modified light source and microscope/lens setup in order to work with
opaque substrates. An important constraint with this approach is the physical size
of the well. The walls of the assay well need to be high enough to contain enough
cell media during experiments. Therefore they will limit how close the microscope
objective lens can get to the sample. A possible solution is to use a microscope that
has a long working distance at the cost of lower magnification, and loss of image
quality due to refractive effects.
5.2.4 Previous Approaches
Imaging cells on top of CMOS chips has been performed using different mi-
croscopy methods, including normal light [10, 11, 37, 57], scanning electron [79, 80,
81, 82], and fluorescence-based methods [83, 84, 85, 86, 87]. Park et al. obtained
fluorescence images of cells at 24 hour intervals to validate impedance measurements
69
[85]. They also monitored cells during extracellular stimulation and recording [83].
Chi et al. used a fluorescence image to validate cellular impedance and optical mea-
surements [84]. Lopez et al. and Abbot et al. used fluorescence images to locate
cells for comparison with extracellular recordings [86, 87]. Niitsu et al. used light
microscopy to correlate cell locations with cyclic voltammetry measurements [37].
Nabovati et al. used light microscopy to image cells on top of a capacitance sensor
chip [11]. Couniot et al. used images of cells to count cell coverage in order to
quantify the sensitivity of their device [10]. However, in each of these cases, imaging
was done at discrete time points outside the incubator using standard imaging plat-
forms. This is sufficient for some applications such as extracellular cell monitoring
where a single snapshot of neuron locations is adequate to localize sources. How-
ever for other applications such as cell proliferation or motility monitoring, these
imaging techniques do not meet the temporal resolution needed to correlate with
cell morphological changes.
5.3 This Work
5.3.1 In-incubator Imaging
This work uses an in-incubator platform to perform long-term imaging of live
cells on LoCMOS. A standard reflective microscope head unit (Microzoom II, Bausch
& Lomb) with a camera (16 MP, Amscope) is secured to a custom-machined mi-
croscope assembly that was designed to fit within the incubator (Isotemp, Fisher
Scientific). A white LED lamp (MR16, EcoSmart) with a peak emission wavelength
70
of 960 nm was used as the light source. The CMOS chip is placed under the mi-
croscope on top of a two-axis linear stage (Thorlabs) to facilitate XY alignment,
with height controlled using the vertical adjustment of the microscope head unit.
A standard objective lens with a long working distance (>2 cm) is used to facili-
tate positioning of the LoCMOS within a cell culture well. The sensor is powered
and automatically collects data while images of the sensor surface are automatically
taken every 5 minutes. Since the system has minimal wiring requirements, it does
not disrupt normal operation of the incubator, ensuring a stable culture environ-
ment (37 ? C, 5 % CO2). This allows data to be recorded over the course of the
experiments (up to 3 days), in an unsupervised and uninterrupted manner.
5.3.2 Dual-Well Approach
A dual-well structure, as shown in Fig. 5.3a, was developed to alleviate the
challenges of imaging CMOS chips. The inner well serves as the primary culture
well where cells interact with sensors. This well is filled with cell media and covered
with a glass coverslip to prevent media evaporation and keep a fixed focal distance
for microscopy. A larger outer well is also filled with cell media but is left uncovered
to allow gas exchange and maintain physiological pH levels. Ports are created in the
inner well to allow for diffusion between the two wells, thereby maintaining proper
growth conditions in the inner well.
The size and shape of the well can be customized for different package sizes.
In this work, the primary well was formed using a cuvette (10?10 mm) that was
71
outer well chip carrier
inner CMOS IC glass 
well
CMOS IC coverslip
diffusion ports cells cell media
(a) (b)
Figure 5.3: a) Schematic of dual-well package. b) Photograph of packaged chip inside an
incubator
glued onto the packaged device using a using a bio-compatible silicone glue (Kwik-
Cast, World Precision Instruments). The cuvette had 5 mm holes drilled through
the bottom of two of its sidewalls to create the diffusion ports. The secondary well
was formed by attaching a larger plastic tube (30 mm diameter) on the device.
The outer well should be taller than the inner well to prevent media overflow. A
photograph of a packaged device is shown in Fig. 5.3b.
5.4 Live Cell Experiments
5.4.1 Experimental Protocol
Human ovarian cancer cell lines A2780 and CP70 were prepared following the
protocol described in Section 4.3. Prior to each experiment, the packaged devices
were sterilized by exposure to UV light for 30 minutes. They were then sequentially
rinsed with deionized water, phosphate buffer solution, and fresh cell media. Both
72
wells on the device were filled with pre-warmed cell media and then 105 cells/ml
were added directly over the sensing area using a micropipette. More media was
then added to fill the primary well and then a coverslip was used to seal the well.
Excess media was then aspirated out from the secondary well, allowing the coverslip
to form a tight seal around the primary well. The device was then placed inside an
incubator and recording was started. Cells were allowed to grow for 72 hours before
the experiment was stopped.
5.4.2 Experimental Results
In order to validate the capacitance sensor measurements, experiments were
first conducted with the imaging and sensor recordings running concurrently. Fig.
5.4 shows time-lapse images taken using the in-incubator imaging system along with
the corresponding capacitance measurements for four sensors in the array. During
this portion of the experiment, cells were pipetted into the culture well over the
sensor array, then allowed to settle onto the chip surface in a random manner.
As cells first seeded onto the surface they were rounded, as can be seen in image
A of Fig. 5.4a. While this initial binding is weak, a noticeable difference was seen in
channels 2 and 4, which had an initial seeding of cells compared to channels 1 and
3, which were void of cells. Over the next several hours, the cells began to flatten
and adhere more strongly, as can be seen in the progression to images B and then C.
Corresponding capacitance increases to 380 aF and 600 aF were seen in channels 2
and 4, respectively. Channel 3 showed a brief increase in capacitance between hours
73
A 1 2 B C D
3 4
E F G H
(a)
800 A B C D E F G H
700
600
500
400
300
200 800
Ch. 1
100 700 Ch. 2
Ch.3
0 600 Ch. 4
-100 500
5 10 15 20 25 30 35 40
Time (hours) 400
(b)
Figure 5.4: Correlation of capacitance measurements and time-lapse imaging of cells on
the chip during cell-substrate binding. a) Images taken of four sensors at different time
points. b) Corresponding sensor responses showing the measured capacitance changes over
time. The letter markers indicate the time points corresponding to each frame. Video is
available as supplementary material.
19 (image D) and 22 (image E). This was due to the cell on the top-right of the
electrode, which traversed from the top of the electrode to the bottom over this time
period, briefly increasing the capacitive coupling. Additionally, channel 1 showed a
sharp increase in capacitance of ?200 aF at 21 hours. This was due to the cell to
the top-right of its electrodes, which moved downwards towards the center of the
electrodes and remained there, as can be seen between images D and E. Channel
74
| C|
C|
2 measured a brief 215 aF dip in capacitance at 19 hours. This was due to the
cell on top of the electrodes (image C) undergoing cell division. The two resulting
daughter cells (image D) covered approximately the same area of electrodes as the
parent, but the daughter cells were probably not as well adhered, resulting in a drop
in measured capacitance. Over the next several hours these cells started to grow
and to bind to the substrate, resulting in the increase in measured capacitance.
Fig. 5.5 shows results from another experiment where cells were allowed to
grow unhindered. The top panel shows time-lapse images of four electrodes taken
during the experiment and the bottom panel shows the corresponding capacitance
data. The dashed lines and letters indicate the time points for each image.
When cells are first seeded into the culture well, they float in solution before
settling down in random locations on the CMOS chip. At this point they are initially
round and are just beginning to adhere (image A). As time progresses the cells form
stronger connections to the surface, flatten in shape and spread outward (images B
and C). Channel 1 in Fig. 5.5a shows little change in capacitance until 32 hours into
the experiment (image E) where most of the electrode area becomes covered with
cells causing a peak of 600 aF. In the next five hours, the measured capacitance drops
by 200 aF as the electrode coverage area reduces due to one of the cells dividing
(image F). In another example, channel 2 shows a drop of 200 aF from 25 to 27 hours
that is also caused by a dividing cell. Image C shows the parent cell and image D
shows the two daughter cells. Channel 3 shows noticeable increase in coverage
between images A and B, with the corresponding increase of 300 aF during this
time. There is continued electrode coverage through 35 hours of growth followed by
75
A B C
1 2
3 4
D E F
(a)
A B C D E F Ch. 1
Ch. 2
800 Ch. 3
Ch. 4
600
400
200
0
5 10 15 20 25 30 35 40 45
Time (hours)
(b)
Figure 5.5: Correlation of capacitance measurements and time-lapse imaging of cells on
the chip during cell-substrate binding. a) Images taken of four sensors at different time
points. b) Corresponding sensor responses showing the measured capacitance changes over
time. The letter markers indicate the time points corresponding to each frame. Video is
available as supplementary material.
a decrease of roughly 200 aF as the cell shifts downwards. Sensor 4 shows examples
of cells undergoing morphological changes. At 12 hours, a pair of cells stretches
across all for fingers of the electrode, resulting in a bump in capacitance of 350 aF
76
| C| (aF)
(image B). Twelve hours later, the electrode area becomes relatively empty leading
to a drop of 250 aF (image C). At 34 hours cells start to expand and adhere well
across the electrodes, bringing the sensed capacitance to a peak of 750 aF (image
E).
Repeat experiments were performed using the two cell lines and showed similar
growth profiles. The maximum mean capacitance change measured by the sensor
for electrodes fully covered by cells was 1 fF. This results in a coverage sensitivity
of roughly 1.1 aF/?m2.
5.5 Cell Segmentation
As a first step in validating capacitance measurements with ground truth im-
ages of cell coverage, we describe image processing techniques to automatically cal-
culate sensor coverage area for each of the collected time-lapse images. It is then
possible to correlate coverage estimates with capacitance measurements. The pro-
cess of using image analysis techniques to detect cell shape and size is called cell
segmentation.
Images of samples containing dissociated cells, or small fragments of tissues on
glass slides (cytopathology samples) are generally more amenable to segmentation
due to well-separated cell membranes and a clear background. On the other hand,
samples that preserve whole tissues (histopathology samples) are much more com-
plex, with cells of various sizes and shapes embedded amongst a noisy background.
These types of samples however are often prepared with well-characterized stains
77
that can help to identify regions of interest such as cell nuclei and membranes [7].
In this work, cells are grown in vitro directly on top of the CMOS chip and
normal reflected light microscopy is performed. As can be seen in the time-lapse
images in Fig. 5.4a, cells appear as translucent, irregularly-shaped features that are
primarily identifiable through the slight contrast differences along their membranes.
It is also important to note that cells are unlabeled, i.e. no dyes or probes are
added to enhance cell contrast. Another factor that exacerbates image processing
challenges are the visible features of the chip itself, especially the underlying metal
layers that are used for signal routing and electrode formation. These features ap-
pear very prominently in reflected light microscopy and easily obfuscate the outlines
of cells that would normally be visible under a uniform background.
5.5.1 Background
Many different segmentation methodologies have been proposed, however these
can generally be classified into a few categories: intensity thresholding, feature detec-
tion, morphological filtering, and region accumulation [88]. Image intensity thresh-
olding is a basic strategy that assumes cells appear consistently darker or lighter
compared to the image background. Thresholding the image using a global or adap-
tive value can then give an initial estimate of the cell shape and location. Edge
detection and linear image filtering are examples of feature detection algorithms ap-
plied to cell images. They leverage prior knowledge of cell features. Morphological
filtering uses mathematical morphology theory to filter images based on basic image
78
shapes, called structuring elements. These are shapes that are the basis for perform-
ing a dilation (enlarging features), erosion (shrinking features), or combinations of
the two operations. If the locations of cell nuclei or cell centers are known, then
one can iteratively grow regions surrounding that initial seed point to form a group
of labeled pixels. Here, we perform cell segmentation using morphological filtering
and thresholding methods.
5.5.2 Algorithm
5.5.2.1 Pre-processing
As mentioned previously, the visible features of the chip itself introduce a
challenge in traditional segmentation approaches. Therefore, all images are first
processed to reduce the intensity of these features. This is performed by converting
images to grayscale and then normalizing all frames using a reference image con-
sisting of a bare electrode. The first step in pre-processing the images is to register
(align) each frame to the reference. This is required because the sensor surface may
shift slightly over time under the microscope. An iterative intensity-based regis-
tration algorithm is used to perform the alignment. The next step is to perform
the normalization using the reference image. As seen in the sample images in Fig.
5.5a, the central and outer regions of cells are brighter and darker compared to the
background, respectively. To emphasize these two regions, the image normalization
is performed in two parallel paths. First a direct subtraction and scaling is done
directly on the frame and reference. Second, both images are complemented before
79
subtracting and scaling. The result of these two flows are images that emphasize the
dark and bright portions of the cells as can be seen in Fig. 5.6b and c, respectively.
5.5.2.2 Cell Detection
These images are then binarized using an adaptive thresholding approach [89].
This results in preliminary binary masks that delineate cell coverage (Fig. 5.6d).
The red and orange portions of the image correspond to the thresholded dark and
bright regions, respectively. Next, morphological image processing is performed to
denoise the mask based on expected models of cell shape and sensor surface features.
The two basic morphological operations are erosions and dilations. The mask images
are probed using a disk-shaped structuring element, to shrink or enlarge the base
image. An erosion is performed using a structural element to filter out any false
positives caused by features of the chip that were not removed during reference
subtraction. This is evident, for example, in the false positives created by the
A B C
D E F
Figure 5.6: Images of intermediate steps in cell segmentation algorithm. a) Original image
of cells over sensor electrode. b, c) Result of registering and subtraction of bare reference
electrode for dark (b) and bright (c) emphasis. d) Output of two-level adaptive threshold-
ing with orange and red regions corresponding to bright and dark levels, respectively. e)
Mask after morphological filtering. f) Detected cell coverage regions and electrode area,
marked in red and orange outlines, respectively.
80
interdigitated electrodes in Fig. 5.6d. Dilation is then performed to expand the mask
area around edges that were not well defined after reference subtraction. The result
of these operations on the bright and dark emphasized masks are then combined
(Fig. 5.6e) and dilated once more. Finally, any remaining holes in the mask are
filled. Fig. 5.6f shows the outline of the final mask (red) along with the outline of
the electrode sensing area (orange). The overlap between the two regions is the final
estimated coverage for each frame. Listing 1 and 2 outlines the image processing
steps. The comment lines include letters in the parentheses that correspond to the
sample images in Fig. 5.6.
Listing 1 Cell segmentation pseudocode
//register image (a)
x = registerIm(input,ref);
xc = complementIm(x);
refc = complementIm(ref);
darkIm = processIm(xc,refc);
if (doBrightIm) {
brightIm = processIm(x,x0)
}
//combine and fill mask (f)
mask = darkIm | brightIm;
mask = dilate(mask, line0/90);
mask = erode(mask, disk2);
mask = fill_holes(mask);
81
Listing 2 Image processing pseudocode
function processIm(x,x0) {
//subtract reference (b,c)
xref = x-x0;
//binarize image (d)
mask = threshold(xref, 3 lvl, adaptive);
//clean mask (e)
mask_hz = erode(mask, horizontal lines);
mask_hz = dilate(mask, line0/90);
mask = mask & ~mask_hz;
mask = erode(mask, disk1);
mask = fill_holes(mask);
return mask;
}
5.5.3 Results & Discussion
The cell coverage detection algorithm was then evaluated using a dataset of
images generated through the imaging and cell culture protocols described in Sec-
tions 5.3.2 and 4.3, respectively. Three experiments were performed using the two
cell lines (1 ? A2780, 2 ? CP70), and a subset of the data consisting of capacitance
measurements from nine sensors and their corresponding time-lapse images (3196
frames) was used for analysis.
5.5.3.1 Sensor Correlation
The image processing was performed on each image in the dataset resulting in
an estimate of cell coverage as a function of time. Fig. 5.7a shows examples of time
series results of the segmentation. The red traces (scale on right axis) are the esti-
mated electrode coverage area and the black traces (left axis) are the corresponding
82
capacitance measurements. Through multi-physics simulations [55] and observation
of the time-lapse images and sensor data, we expect capacitance measurements to be
proportional to the area coverage of the electrodes. Therefore, if a larger area of the
electrodes is covered we expect a larger corresponding change in capacitance. As can
be seen in Fig. 5.7a, there is indeed a clear qualitative correlation between coverage
and capacitance. At the beginning of experiments when cells have just settled, they
are at their smallest size leading to a low electrode area coverage and at the same
time the capacitance is relatively low. In both examples, capacitance increases after
10 hours and there is a corresponding increase in coverage. Interestingly, decreases
in capacitance are also correlated between the two curves. For example, the drop
in capacitance between 21 and 23 hours for dataset 1 is reflected in the coverage
estimate as well. It is important to note, however, that the capacitance and cover-
age scales are not expected to have one-to-one correspondence. Indeed, one expects
the capacitance gain (capacitance per unit area) would vary from pixel to pixel due
to process variations. To better quantify the capacitance-coverage relationship, we
look at scatter plots of capacitance vs. coverage (Fig. 5.7b). As can be seen, there
is a roughly linear relationship between the two variables (rcov = 0.93 for both,
Pearsons correlation coefficient). The majority of the data points follow the linear
trend although several clusters of outliers exist. For example, the cluster of points in
dataset 1 that lie below the best-fit line at 20?30 hours is due to the over-reporting
of the electrode coverage by the detection algorithm. The correlation coefficients
have a mean of 0.80 and standard deviation of 0.12, across the nine datasets, each
reflecting data collected from a single sensor over three experiments.
83
a 1000 40 750 c
750 30
500 50020
250 10 250
0 0 0
0 10 20 30 40 0 10 20 30 40 -500 -250 0 250 500
Time (hours) Time (hours) Residuals (aF)
b 1000 2.5
2.0 d750
500 1.5
250 1.0
0 0.50
0 10 20 30 40 0 10 20 30 40 0.5 0.6 0.7 0.8 0.9 1.0
Estimated Coverage (%) Estimated Coverage (%) Correlation Coeff.
Figure 5.7: Correlation of measured capacitance and estimated cell coverage. a) Capac-
itance and coverage plotted as a function of time for two exemplar datasets. b) Scatter
plots of capacitance vs. estimated coverage, with linear regression line, for the data in
(a). c) Histogram of the difference between fit line and measured capacitance values for
all datasets (N = 3196). d) Derived sensor gain vs. correlation coefficient.
A histogram showing the residuals between measured capacitance values and
the linear best-fit lines is shown in Fig. 5.7c. The figure reflects data from all 3196
frames/datapoints. A Gaussian fit of the data (R2 = 0.97) results in a distribution
with a mean of ?1 aF and standard deviation of 80 aF. Using a 2? threshold to
cover 95 % of the distribution, the error would be 1? 160 aF.
Several factors contribute to this relatively large error, with variability result-
ing from both the segmentation algorithm and also biological responses. First, the
quoted error metric is calculated under the assumption that the coverage estimates
from image processing provide ground-truth data, and that deviations from the best-
fit linear line are the consequence of error in capacitance measurements. A more
accurate figure could be obtained if each of the images in the dataset were manually
processed to mark areas of cell coverage. This ground truth dataset could then be
used to refine both sensor calibration and algorithms for estimating cell coverage.
84
| C| (aF) | C| (aF)
Est. Cov. (%)
Gain (aF/ m2) Count
Another source of variability is the quality of cell binding for given coverage
area. Cells that are in their initial state of attachment (e.g. cells that have just been
seeded into a well) primarily stick to the substrate through electrostatic interactions
[26]. Over time, cells form a complex structure of cell adhesion molecules that bind
them to the substrate, thus increasing the coupling strength. Therefore, given two
cells that cover the same surface, the capacitive coupling of the two will be different
depending on how well developed the bonding is. Since image processing treats area
coverage equally, this binding quality is not considered.
5.5.3.2 Sensor Gain
Given the expected linear relationship between electrode coverage and capac-
itance, the slopes of the best-fit correlation lines calculated in Section 5.5.3.1) (and
shown in Fig. 5.7b) are a measure of the sensors gain. Fig. 5.7d shows a scatter plot
of the slope values as a function of rcov for each of the nine datasets. The slopes
range from 0.32 to 2.00 aF/?m2 with a mean of 1.11 ? 0.52 aF/?m2. These val-
ues are of course highly dependent on the accuracy of coverage estimation. Indeed,
datasets with higher rcov show a tighter distribution. If we focus on datasets with
a good capacitance-coverage correlation (rcov > 0.85), we see a mean sensor gain
of 1.28 ? 0.29 aF/?m2. This value is close to the initial estimate of 1.1 aF/?m2
obtained using the mean maximum measured capacitance (Section 4.4). It is also
with range of the sensitivity values in existing literature (see Table 8.1). The result
can also be improved given a larger number of image and measurement datasets.
85
Further improvements of the cell segmentation algorithm can also be performed by
leveraging machine learning or neural network structures to improve classification
accuracy.
86
Chapter 6
Applications, Part 1: Mitotic Activity Detection
6.1 Preface
Here were present work towards using the capacitance sensor chip as a tool for
cancer research applications: detecting cell mitotic activity. This is relevant in two
respects: 1) mitotic activity in tissue is an important indicator in cancer prognosis,
and 2) detection of mitosis events can be used as a marker for cell cycle-targeted
therapy development.
This chapter includes excerpts of a manuscript in preparation with authors
including myself, Sheung Lu, Dr. Marc Dandin, Dr. John Basile, Dr. Elisabeth
Smela, and Dr. Pamela Abshire. My contributions include setting up and perform-
ing the experiments, image and data analysis for mitosis event detection, developing
the automated detection algorithm, and most of the writing.
6.2 Introduction
6.2.1 Mitotic Counts
Histologic grading, the visual classification of cancer cells and tissues, is an
important prognostic factor and indicator of metastatic risk in cancer patients.
Counting cell divisions in particular is an important feature that has been linked
87
to grading. For example, Andreou et al. looked at primary, localized, high-grade
soft tissue sarcomas and determined that mitotic count (along with the amount
of viable tumour remaining post-treatment) strongly correlated with prognosis and
could serve as the basis for designing personalized treatments [90].
Pathologists currently obtain mitotic counts manually by optical inspection
through a high-powered microscope. Tissue samples are first affixed and stained.
Then the number of cells undergoing mitosis is manually counted by identifying
morphological features that match the latter part of the cell cycle (e.g. visible
chromatid separation, lack of nuclear membrane, hyperchromatic appearance) [91].
Unfortunately, counting mitoses can be subjective and affected by fixation of the
tumour tissues, the quality of the section or biopsy, and adequate selection of the
most representative areas [92]. As such, it is a tedious and time-consuming task,
and has limited reproducibility. Rapid counting also results in errors so multiple
recounts are generally required [93]. Therefore a method of objectively quantifying
cell proliferation and mitotic events, can be a great benefit for cancer prognosis.
Quantitative image analysis techniques have been developed to address some
Figure 6.1: Example images of nuclei that are challenging to detect (figure taken directly
from [7])
88
of these challenges. These systems analyse large datasets of stained tissue samples
and perform mitosis detection through texture or colour-based feature analysis [94,
95, 96]. More recently, machine learning has been leveraged to undertake this task.
They include conventional neural network methods that use manually-engineered
feature descriptors for classification [97, 98], along with deep neural networks that
are feature-agnostic [99, 100]. However, variations in images during sample staining,
the large amount of visible objects, and the similar appearance of mitotic nuclei
with non-mitotic structures, make image classification a difficult task [101]. Fig.
6.1 shows examples of nuclei that are challenging to detect.
6.2.2 The Cell Cycle
A second important application for mitosis detection is in the development of
targeted chemotherapy regimens. Many modern chemotherapeutic regiments tar-
geting cancer cells focus on disrupting the cell cycle, which is the process by which
cells grow and multiply. The cell cycle is divided into different phases that mark cell
transitions from resting behavior, to performing DNA synthesis and chromosome
separation, and finally division [102]. Checkpoints exist within the cell cycle to en-
sure only healthy cells proliferate, while detection of defects will trigger a cell cycle
arrest. This ensures that any flaws are not passed on to daughter cells when cells
undergo division. Disruptions to the progression of this cycle that allow unchecked
cell multiplication is a hallmark of cancer cells [103]. As such, modern advances in
therapeutics have focused on targeting the cell cycle in cancer cells. Specifically,
89
anti-cancer drugs have been developed that focus on inhibiting cyclin-dependent ki-
nases (CDKs), which are enzymes required for driving the cell forward through the
different phases of the cycle. Furthermore targeting CDKs during certain phases of
the cell cycle has shown promise in improving the efficacy of cancer treatment at
low dosages that limit non-specific damaging effects on normal cells [103, 104, 105].
Therefore by careful administration of the drug during known phases of the cell
cycle, one may be able to improve the efficacy of the drug treatment while reducing
required dosages. A real-time cell viability monitor that can identify points in the
cell cycle could thus be a useful tool in the screening process.
6.3 Detecting Morphology Changes
Here we propose an alternative approach to detect mitotic events: monitor
changes in cell morphology through cell-substrate binding measurements. The basis
for this approach is due to the acute changes in cell shape and size that occur during
the final phase of the cell cycle, cytokinesis, when a cell divides. The parent cells
membrane contracts, causing the cell to round and decrease its footprint over the
substrate. A contractile furrow forms and pinches at the cells equator, eventually
dividing the cell into two daughter cells [102]. These new cells then begin adhering
and spreading out onto the substrate. These changes in morphology affect the ca-
pacitive coupling at the cell-substrate interface which can then be detected through
capacitance measurement.
90
6.4 Mitosis detection algorithms
Two mitosis event detection algorithms were compared, one based on a moving
average indicator, and the other based on template matching.
6.4.1 Kaufman Adaptive Moving Average
The Kaufman adaptive moving average (KAMA) algorithm [106], originally
developed as a financial indicator, is used to generate a capacitance indicator curve.
For each data point xi, we find the absolute difference, di, with the capacitance
value W points prior (xi?W ), where W is the length of the averaging window, as
given by Equation 6.1. The volatility of measurements within this window, vi, is
computed by summing the difference between successive data points in the window,
as given by Equation 6.2.
di = |xi ? xi?W | (6.1)
?W
vi = (|xi?n+1 ? xi?n|) (6.2)
n=1
The ratio of di and vi is then scaled by parameters ? and ?, to generate a
smoothing constant SCi, as given by Equation 6.3. The smoothing constant is then
used to scale the difference between the raw capacitance value and prior KAMA
value, as given by Equation 6.4. The result is an exponentially weighted moving
average indicator (MAI) that smooths over transient fluctuations in capacitance and
tracks behind larger swings. This means that a crossover of the raw sensor output
with the KAMA line (i.e. the data crosses below the KAMA line) is indicative of
91
a directional shift in the data. It follows that the opposite crossover (i.e. the data
crosses back above the KAMA line) then indicates an end to the shift. This feature
can be used to mark a mitosis event because during the division, the cells contract
and are expected to cause a drop in capacitance as surface area coverage is reduced.
This is followed by a recovery phase where the two new daughter cells re-adhere
to the substrate. The time taken for a capacitance drop and recovery is dependent
on cell morphology kinetics, therefore events detected through the KAMA-signal
crossover points are thresholded using a maximum time, Tmax, to filter out false
positives.
di
SCi = ( (?? ?) + ?)2 (6.3)
vi
yi = yi?1 + SCi(xi ? yi?1) (6.4)
6.4.2 Template Matching
Template matching (TPM) algorithms compare the entire recorded signal with
a template that is representative of the signal being detected. The metric used
to quantify this similarity can be computed in several different ways. Common
measures include the Euclidian-distance, cross-correlation coefficient, and the sum
of squared difference (SSD). Here the SSD is used to compute the similarity index
between raw signal and template, as given by Equation 6.5.
?N
SSDj = (xj,i ? ti)2 (6.5)
i=1
92
where t is the template with length N , and xj is the jth window of the raw data
being compared to the template. If the template matches the window of the raw
signal identically, then the SSD value will be zero; slight differences between data
points will shift the SSD upwards. Therefore, after computation of the SSD metric,
a threshold is applied to identify detected events. In this dataset, the template was
created using the time-aligned median of five sample marked events.
6.5 Experimental Results
6.5.1 Mitotic Activity Detection
Fig. 6.2a shows examples of cells dividing at a single sensors electrodes. The
bottom panel shows the measured capacitance over a period of 60 hours. The top
panel shows three pairs of images of the sensor and surrounding area at specific
time points indicated by the square and circle markers. In this example, three
occurrences of division took place at the same sensor, each of which are highlighted
in the capacitance time-series data. The square and circle symbols mark when
images were taken and show cells before and after division occurs, respectively.
During each division the parent cell begins to contract, reducing its surface area in
contact with the electrode and causing a reduction in cell-substrate coupling. This
effect can be seen in the sensor readings as the downtrend of > 50 aF that occurs
over a period of roughly 60 minutes. The cell then divides into daughter cells,
which then expand and settle over a larger area of the electrodes. This process in
turn increases the measured capacitance. Fig. 6.2b shows manually verified mitosis
93
events for a single experiment for multiple sensors.
Analysis of the detected events can be performed to elicit cell division kinetics.
600
200
500
100
400 0
0 50 100 150
Tdrop (min.)
300
Tdrop
200
Trise
100
0
0 10 20 30 40 50 60
Time (hours)
(a)
(b)
Figure 6.2: a) Top: Images taken of a sensor at different time points. Bottom: Cor-
responding sensor response showing the measured capacitance changes over time. The
highlighted portions indicate division events marked by the detection algorithm. Inset:
Distribution of the time taken from start of division event to reach minimum capacitance
(Tdrop) and time taken to recover (Trise). N = 41 detected events. b) Examples of detec-
tion mitosis events from one experiment. Capacitance traces have been offset for plotting
purposes. Video is available as supplementary material.
94
| C| (aF)
Trise (min.)
Using the measured data we extract the time taken from the detected start of an
event to the time at which a minimum capacitance is detected within the events
period (Tdrop). This indicates the portion of the division cycle where the contractile
ring causes the cell to shrink and round. The corresponding recovering time (Trise)
is the time for the daughter cells to adhere and re-spread causing a rebound in
sensed capacitance. The inset of Fig. 6.2a shows a scatter plot comparing Tdrop and
Trise. Each point corresponds to a division event that was automatically marked
using the detection algorithm, and manually verified with time-lapse imaging (N =
41). The dotted line has a slope of 1, as a reference. The majority of the drop
and recovery phases range between 20 and 100 minutes but interestingly show a
roughly linear time correspondence with the recovery phase taken slightly longer
than the corresponding drop phase. In fact, one might expect the recovery phase
to take significantly longer since daughter cells are rounded immediately following
mitosis and have to re-establish cell-substrate adhesion, a situation similar to cells
making initial attachments to their substrate. However, recent research in cell-
substrate adhesion during mitosis has revealed the existence of remnant adhesion
molecules that aid in post-mitotic daughter cells to re-spread and anchor to the
substrate[107, 108]. This may help explain the relatively quick recovery of the
measured binding capacitance.
Each sensor electrode covers an area of 30 ? 30 ?m2 which means a total
sensing area of 0.0144 mm2 across the array. A total of 41 manually verified mitotic
events were detected across four experiments, resulting in a mean mitotic figure of
710 divisions/mm2 across the four experiments. This figure is significantly higher
95
than typical counts of < 15 divisions/mm2 [92, 109], primarily because typical
image-based counting methods are based on analysing tissue samples extracted at
a fixed point in time. Here the count index reflects divisions that occur across a
period of three days and thus show a fuller picture of cancer cell activity. If the
time period of 72 hours is taken into consideration, the scaled figure would be 9.9
divisions/mm2/hr, which is within the typical range for histological grading.
6.5.2 Detection Algorithm Performance
In order to develop the sensor as a tool for counting mitotic figures, we eval-
uate the two detection algorithms for specificity and sensitivity. Receiver operating
characteristic (ROC) curves were generated for each algorithm using the manually
marked dataset as ground-truth. For the TPM algorithm, the SSD threshold was
varied to generate the true positive rate (TPR) and false positive rate (FPR) of the
detected events. For the MAI algorithm, the maximum time span between crossover
points (Tmax) is varied to generate the ROC points. Fig. 6.3 shows ROC curves
for the two algorithms. The dotted line indicates the random guess boundary; bet-
ter performing algorithms would be in the upper half of the boundary and ideally
following along the y-axis (high TPR and low FPR) before saturating near a TPR
of 1. At stringent threshold levels, the TPM algorithm outperforms MAI with a
consistently better TPR. When the threshold is relaxed to allow more detections,
MAI begins to outperform TPM, starting at an FPR of roughly 0.25.
96
1.0
0.9
0.8
0.7
0.6
0.5
0.4 TPM
0.3 MAI
Random guess
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
False Positive Rate
Figure 6.3: Receiver operating characteristic curve for cell detection algorithm. The two
lines correspond to varied minimum capacitance changes and durations for mitotic events.
97
True Positive Rate
Chapter 7
Applications, Part 2: Drug Screening on Chip
7.1 Preface
Here we presnt work for using the capacitance sensor chip as a proof-of-concept
tool for monitoring chemotherapeutic agent potency in cancer cells. Drug-sensitive
and drug-resistive cells are cultured on the LoCMOS device and are exposed to
cisplatin, a platinum-based cancer drug. Results show the ability to discriminate
cell lines using capacitance sensor measurements. The data is then used to extract
cell death kinetics.
This chapter includes excerpts of a manuscript that has been submitted to
Biosensors and Bioelectronics, an article that was authored by myself, Sheung Lu,
Dr. Marc Dandin, Dr. John Basile, Dr. Elisabeth Smela, and Dr. Pamela Ab-
shire. My contributions include setting up and performing the experiments, data
processing, analysis of cell death kinetics, and most of the writing.
7.2 Cell-Based Drug Screening Assays
Cell-based assays are a vital tool in the drug discovery process and can be
used to quantify the cytotoxic effects of drugs on cells and identify new targets
for drug development. As mentioned previously, most current cell-based screening
98
studies require the use of specialized laboratory equipment and reagents, and provide
data through methods that require labeling such as fluorescent, bioluminescent, and
colorometric methods. While these methods can provide high-throughput analysis
through parallelization of experiments, they generally have high operational costs,
can be labor intensive, require sub-sampling of analytes, and may lack specificity
[110]. Furthermore most conventional methods provide data at sparse sampling
points or are endpoint assays, and so do not provide information on the dynamics
of cell growth.
Real-time monitoring of cell viability over the course of drug exposure and
subsequent cell death may provide a more complete picture of the cytotoxic effects
of drugs, enabling the development of a more precise drug dosage regimen. Indeed
modern chemotherapeutic regimes generally involve the use of multiple drugs [111]
administered sequentially or in an alternating fashion, with the aim of reducing
drug resistance, reducing non-specific toxic effects, and maximizing overall efficacy.
Computational models have also been developed to generate adaptive therapeutic
[112] regimes or select optimum strategies to address issues of tumor heterogeneity
[113]. Therefore the ability to perform continuous cell measurements in real-time
would represent a great advance over current methods of determining drug cytotox-
icity and cancer cell susceptibility, replacing current laboratory assays with a more
rapid and more relevant alternative that could hasten the drug development process
toward clinical implementation.
Most chemotherapeutic agents are designed to induce cell death through a
process called apoptosis, or programmed cell death. This is a regulated process that
99
causes the nucleus to condense, organelles to pack together, and cell to shrink in
overall area [20]. This change in morphology is thus an attractive candidate for
observation through real-time cell-substrate capacitance measurements.
7.3 Cell Response to Chemotherapeutic Agents
Cisplatin, a widely used platinum-based chemotherapeutic drug, was used as
the anti-cancer agent. The agent was obtained in dry powder form (BioVision) and
prepared with DI water to create a 1 mM stock solution. The solution was stored
at room temperature away from light. Human ovarian cancer cell lines A2780 and
CP70 were prepared following the protocol described in Section 4.3. For treat-
ment, cells were plated into the device and allowed to grow for 48 hours before the
chemotherapeutic agent was administered directly into the cell growth well.
The top panel of Fig. 7.1 shows exemplar images of cells in the vicinity of
two sensors through a drug administration experiment. The left-most image, taken
at t = 26.8 hours shows cells growing normally. The 100 ?M dose of cisplatin was
administered at t = 53.5 hours, indicated by the red line, and the second image shows
the surface 6 hours later. Cells still appear healthy although a downtrend in the
measured binding capacitance can be observed in the sensor data traces. Roughly 4
hours later, a more noticeable change in morphology of the cells can be seen, both
in the images and data. The bottom panel of Fig. 7.1 shows the corresponding
data from each of the two sensors. Capacitance decreases of 230 aF and 440 aF are
measured in channels 1 and 2, respectively. Images of the chip surface show that the
100
A 1 2 B C
1000
Ch. 1 A B C
800 Ch. 2
600
400
200
0
0 10 20 30 40 50 60
Time (hours)
Figure 7.1: Top: Microscope images of the two sensors during a drug administration
experiment. Bottom: Corresponding sensor measured capacitance changes over time.
The cisplatin was added at t = 53.5 hours, indicated by the red line and arrow. The labels
on the images correspond to the time points marked by the dashed vertical lines. Video
is available as supplementary material.
cells appear noticeably shrunken and rounded, which is indicative of cell death [20].
This experiment shows the devices ability to track cell health in real-time through
a drug-screening experiment.
We then investigated dose-dependent effects of cisplatin on the two cell lines.
Each cell type was allowed to grow for at least 48 hours, after which different con-
centrations of cisplatin were applied. Control experiments were performed in which
both cell lines were plated and allowed to proliferate unencumbered over the ex-
perimental period. Fig. 7.2 shows the time-series capacitance measurements of the
device as a function of drug dose. Each trace shows the mean response of those sen-
sors in the array that were visually observed to have cell coverage. Sensors that were
void of cells, due to the random initial seeding, were omitted. The red and black
lines correspond to the cisplatin-sensitive (A2780) and cisplatin-resistive (CP70) cell
101
| C| (aF)
0 M 1 M 5 M
1000
A2780
750 CP70
500
250
0
0 25 50 75 0 25 50 75 0 25 50 75
10 M 25 M 100 M
1000
750
500
250
0
0 25 50 75 0 25 50 75 0 25 50 75
Time (hours)
Figure 7.2: Mean capacitance sensor responses for various concentrations of applied drug.
Each chart shows the response for two cell types, A2780 (drug-sensitive), and CP70 (drug-
resistive). The dashed lines indicate when the drug was applied to the cell culture. In the
control experiment both cell lines were viable through the experiment. Addition of 1 ?M
cisplatin showed an immediate effect on the capacitive coupling of A2780 cells while CP70
cells remained stable. Higher concentrations began to induce a response in CP70 cells as
well.
lines, respectively. The dashed black lines indicate the time at which the drug was
administered.
The time-series data shows how drug dosage affected cell viability. As ex-
pected, the control experiment showed a continual increase in measured capacitance
as cells proliferated without damage. At the same time, for high dosages, both cell
lines were observed to be adversely affect. As drug concentration was reduced from
100 ?M , the net effect diminished. At a concentration of 1 ?M , the resistant CP70
line was seen to be relatively unaffected by the drug whereas the sensitive A2780
102
| C (a)| | C (a)|
line showed a downtrend. This observation was confirmed by imaging the surface
after the experiment was completed. The difference between the two responses can
be highlighted by looking at the net change in measured capacitance from the time
of drug application and 24 hours later. The resulting chart is shown in Fig. 7.3.
There was a net positive change in signal for the CP70 line for drug concentrations
up to 5 ?M , although the magnitude decreased with increasing concentration. At
the same time, the A2780 line showed an immediate decrease in viability with a 1
?M cisplatin dosage. At a concentration of 10 M, both cell lines were affected, as
mentioned previously.
600
A2780
400 CP70
200
0
-200
-400
-600
-800
0 1 5 10 25 100
Drug Concentration ( M)
Figure 7.3: Measured difference in average capacitance from immediately after the drug
was administered to 24 hours afterwards.
103
Cend - C0 (aF)
7.4 Cell Death Kinetics
We analyze the high-resolution data provided by the device to profile the
temporal response of the cancer cells to cisplatin. The kinetics of cell death are
characterized by the time delay between exposure to the drug and the onset of cell
death (tD), and the rate at which cell death occurs (rD). The onset of cell death was
extracted from the capacitance measurements by finding the time at which the peak
capacitance value (CD) occurs, after the application of the drug. The capacitance
curves were then normalized to CD and the data was fitted to an exponential decay
model:
?C(t)
C(t) = = C0 exp(?rDt) + CB (7.1)
CD
where t (hr) is time, rD is the decay time constant (hr
?1), C0 and CB are initial
and baseline values, and C(t) is the normalized capacitance as a function of time.
Cells were cultured as described previously and exposed to 25 and 100 ?M
dosages of cisplatin, concentrations that were previously confirmed to be lethal to
both cell lines. Fig. 7.4a shows the normalized capacitance measurements of two
sensors from experiments where A2780 and CP70 cell lines were exposed to 25 ?M of
cisplatin. For clarity, the measurements from the start of cell culture up to the time
at which drug is administered are omitted. Additionally, the data is time-aligned
such that the onset of cell death occurs at t = 0 hrs. Therefore tD is the time span
between the first data point and t = 0 hrs, as indicated by the horizontal arrows.
Fig. 7.4b shows the corresponding plots for a dose level of 100 ?M . The dashed
104
lines in both figures are the fitted exponential decay models for each trace. Fig. 7.4c
shows the mean tD for each dosage and cell line. As can be seen, there is a significant
decrease in the time to onset of cell death when drug dosage is increased from 25 to
100 ?M (p = 0.003 and p = 0.004 for A2780 and CP70 cells, respectively). Fig. 7.4d
shows the mean rD values. No significant difference in the decay time constant was
seen as a function of drug concentration (p > 0.05). Interestingly, this parameter is
higher for the resistant CP70 cell line compared to the sensitive A2780 line.
Drug delivery in these experiments involved dispensing microliter volumes of
the drug into a relatively large well, so drug exposure relied primarily on diffusion
1.0 a) tD 1.0 b)
0.8 0.8
0.6 0.6
0.4 0.4
t
0.2 D 0.2
0 0
-10 0 10 20 30 -10 0 10 20 30
Time (hr)
15.0 **c) 0.25 d)
12.5 ** 0.20
10.0 0.15
7.5
5.0 0.10
2.5 0.05
0 0
25 M 100 M 25 M 100 M
Drug Concentration ( M)
Figure 7.4: Normalized capacitance values vs. time for experiments where cells were
exposed to drug concentrations of a) 25 ?M and b) 100 ?M . The bottom panel shows
the mean extracted a) time delay and d) decay time constant for the set of experiments.
(? ? p < 0.01)
105
Time Delay (hr) C/CD
Decay Rate (hr-1)
through the cell media. As mentioned previously, microfluidics could be incorpo-
rated into future versions of the device in order to better control dosage. With
channels running across the sensing area, one could precisely control flow rates to
maintain a constant drug level. Additionally, exposure times could be controlled,
allowing for evaluating different dosage regimens with multiple drug combinations.
Furthermore, drug dispensing could be regulated by incorporating cell cycle marker
detection, creating a feedback control loop for automated cell cycle-targeted drug
screening therapies.
7.5 Summary
This chapter presented results of in vitro experiments to monitor chemothera-
peutic agent potency in cancer cells in real time, using a CMOS capacitance biosen-
sor. Experiments were performed to examine the effects of a chemotherapeutic
agent on drug-sensitive and drug-resistant human ovarian cancer cell lines. The ca-
pacitance measurements were used to extract parameters of cell death kinetics and
showed changes in the time to the onset of cell death as a function of drug concen-
tration. Results additionally showed successful discrimination of the sensitive and
resistant cell lines, and confirm the biosensor as a proof-of-concept tool that has
potential to hasten the drug development process.
106
Chapter 8
Conclusions
8.1 Summary
In this dissertation, we presented a CMOS-based capacitance sensor for on-chip
cell monitoring with high-resolution measurement capabilities to perform single cell
analysis. The sensor was specifically designed with simplicity, robustness, and ease-
of-use in mind. The chip uses the capacitance-to-frequency based sensing paradigm
that alleviates the mismatch and baseline offset issues that affect other approaches,
as discussed in Section 2.2.5. A more thorough validation of the sensor was then
explored. We developed a platform to perform time-lapse microscopy of cells on
the sensor surface over multiple days of experiments. The high-resolution images
that the system provides, allows for validation of sensor responses and confirmed
the sensor?s capability to detect changes in individual cell morphology. Finally, we
performed proof-of-concept drug screening experiments as a first step in validat-
ing the system for use in real-world applications. to characterize the potency of
chemotherapeutic drugs on cells.
This chapter outlines some of the outcomes that this work has produced to ad-
vance the field of biological cell monitoring forward. A comparison of state of the art
is presented, discussing traditional laboratory techniques, lab-on-CMOS approaches,
and the improvements made in this research. A discussion of future extensions and
107
applications is then presented.
8.2 Comparison to State of the Art
8.2.1 Cell-based Assays
As discussed in Section 2.2.2, there are a number of methods currently used in
industry for monitoring cells. These techniques are primarily based on optical de-
tection methods and generally rely on labeling cells with a marker that is indicative
of a characteristic of the cells. For instance Alamar Blue, a commonly-used reagent,
contains a non-fluorescent compound called resazurin that is reduced through the
metabolic activity of healthy cells to generate a fluorescent compound called re-
sorufin [114]. Cells are first incubated with the reagent for 1 to 4 hours and then
a spectrophotometer is used to quantify the fluorescence intensity, which is then
correlated to cell viability [115]. A diagram showing the assay procedure is shown
in Fig. 8.1.
Figure 8.1: Alamar Blue assay procedure. Figure from [8]
108
While these types of methods can be performed in a high-throughput manner
using multi-well plates, the quantification step can only be performed at discrete
time points (generally at the end of an experiment ? an endpoint assay) and so do
not provide high temporal resolution data. This eliminates the opportunity to track
fast changes and to determine temporal correlation between measurements. Further-
more, performing the assays themselves can be labor-intensive, time-consuming, and
require access to large laboratory equipment. Therefore it is important to develop
new techniques to perform these measurements.
8.2.2 CMOS Capacitance Sensors
Several CMOS capacitance sensors have been developed over the years for
monitoring cells that have been tested in vitro. Prakash et al. implemented CS
and differential CBCM designs for characterizing cell adhesion [18, 19, 9, 43, 45]
and presented measurements of muscle cells and cancer cell lines. Ghafar-Zadeh
et al. used a CBCM structure to monitor the growth of bacteria in medium [44].
Nabovati et al. used a differential CBCM structure for cell monitoring [25], extended
the design to an 8 ? 8 array [12], and more recently performed drug cytotoxicity
experiments on chip [11]. Laborde et al. developed a 256 ? 256 ?nanocapacitor?
array using the CBCM approach that was able to track changes in cell morphology
in real-time [13, 116]. Couniot et al. developed a 16 ? 16 array of CBCM sensors
and a multi-electrode CTF oscillator sensor, both used for detecting bacterial cells
[10, 48]. Section 2.2.5 presented discussion of trade-offs of the different architectures
109
used in these devices. Here we discuss the outcomes of this work with respect to
these existing systems.
a) b)
c) d)
e)
Figure 8.2: Example plots of data from CMOS-based capacitance sensors used for in vitro
experiments. Images taken from a) Prakash et al. 2008 [9], b) Couniot et al. 2015 [10], c)
Nabovati et al. 2018 [11], d) Nabovati et al. 2017 [12], e) Laborde et al. 2015 [13]
110
8.2.3 Capacitance Resolution
One of the key challenges in monitoring cells using a lab-on-CMOS platform
is the sub-femtoFarad resolution required to detect changes in the cell-substrate
capacitance. And as shown in this work, single cell analysis requires even higher
resolution ? <100 aF to detection mitosis events, for example. Several of the
aforementioned research groups have presented in vitro measurements of the sensors,
as can be seen in Fig. 8.2.
Panel (a) (Prakash et al.) shows capacitance measurements from two sensors
during an experiment with MDA-MB-231 breast cancer cells. Cells were allowed to
proliferate over a period of 17 hours and then were detached from the sensor surface
(trypsinization) [9]. As can be seen there is a marked response once cells start to
adhere well to the surface of the chip and a corresponding response when they are
detached. However, the signal fluctuates by ?5 fF which can make it difficult to see
single cell events.
Panel (b) (Couniot et al.) shows measurements of three sensors as a 10 ?L
drop of solution containing S. epidermidis bacterial cells is pipetted onto the chip at
t = 20 min. As can be seen there are different time evolutions of the three signals,
and these are attributed to different numbers of bacterial cells at each location. The
authors were able to quantify sensitivity using microscope images and obtained a
figure of 2.2 mV/bacterium. However, due to reported baseline noise levels of 5 mV,
the resolution of the sensor is limited to 450 aF. One thing to note is that the target
application specifically involved bacteria detection, as opposed to animal cells which
111
are > 10 times larger in size.
Panels (c) and (d) (Nabovati et al.) show mean sensor responses in a cell
proliferation experiment (d) and cytotoxicity assay (c). The proliferation experiment
showed an increase in output voltage that was positively correlated with fluorescense
measurements, along with corresponding negative controls. The cytotoxicity assay
involved application of Geneticin, which is a protein synthesis blocking compound,
to a Geneticin-resistant (Hek293) and sensitive (H1299) cell line. Reported results
showed decreases in the growth rate of H1299 cells, but no significant effects on
Hek293 cells. As can be seen the overall population changes can be tracked, but
no results were reported on individual sensors which would be required for higher
resolution, single-cell analysis.
Panel (e) (Laborde et al.) shows the capacitance measurements from a single
sensor throughout a cell proliferation experiment, along with a spatial map of re-
sponses from the entire 256 ? 256 array at certain time points. Given the ? 1?m
spatial resolution of the sensors, the authors were able to monitor a single cell from
multiple sites with a reported 1 aF capacitance resolution. These results are indeed
the best reported figures for CMOS capacitance sensors, and it would be exciting to
see additional in vitro experiments reported with the device. The main drawback,
however, are the complex series of post-processing steps needed to realize the input
electrodes array, and the hardware required for readout (discussed in Section 8.2.4).
In this work we have shown the sensor?s ability to not only track global cell
proliferation, but to also detect single-cell binding events, track cell morphology
changes, identify cell division events, and monitor cell motility. These observations
112
were corroborated with time-lapse ground-truth images of the sensor surface, which
has not been previously performed.
8.2.4 System Complexity
One of the major focuses of this work is to develop a CMOS sensor with a
simple interface and minimal readout requirements. The minimal electrical con-
nection requirements are needed to facilitate packaging the CMOS chips for fluidic
interfaces, which itself is a non-trivial challenge. The number of I/O connections
required depends on several factors including external bias requirements, signals
for chip control, and the readout interface. In this work we reduced the external
I/O requirements to allow full chip control using eight bondpads, which included
a test pad. This figure could be iteratively reduced by removing the test pad, im-
plementing an on-chip clock generator, and simplifying the serial readout protocol.
As a comparison, the other prior work had no strict focus on minimizing wiring
requirements. Instead they relied on wire-bonding, and passivation using an epoxy
[11, 10] or polydimethylsiloxane (PDMS) [13, 116] to provide adequate packaging
yield. While this may be sufficient to perform preliminary biological experiments,
reducing the number of possible failure points in packaging by minimizing electrical
connections can be important for mass production applications.
Another important consideration is signal digitization. For traditional CS or
CBCM sensors, the output is an analog voltage and so must either be buffered off-
chip to an external ADC or converted on-chip. The work presented in [43] and [10]
113
used off-chip ADCs, which simplified the design but provides a path for the addition
of noise into the analog signals. The work presented in [11] and [13] implemented
on-chip ?? and cyclic ADCs, respectively. In these systems external hardware
(FPGA, digital pulse generator) was required to provide required clocking signals
and array readout controls.
In this work, the CTF mechanism is inherently suited for digitization because
the output signal is in a digital format. An on-chip counter is used for digitiza-
tion along with a two-wire serial readout which greatly reduces system complex-
ity and wiring requirements. We also presented system-level architecture improve-
ments to provide a more portable and flexible biosensor by implementing: 1) a
customizable integration time to allow for sensing of a large range of analytes, 2)
programmable I2C addresses that allows multiple peripherals to run concurrently
on the same bus, and 3) spatial scanning of the array controlled with selectable
activation/deactivation of individual pixels.
Finally, post-processing requirements are also an important factor in system
complexity. In [10] the sensor electrodes were void of the foundry-provided passi-
vation layer, and a thinner 25 nm layer of aluminium oxide was deposited using
plasma-enhanced atomic layer deposition. In [12], a thin layer of PDMS was spin-
coated onto the chip surface to act as the electrode passivation layer. In [11], a
polyelectrolyte multilayer (PEM) film was built up through several incubation and
rinsing steps. In [13, 116], the sensor sites themselves are single vias (from the
standard CMOS process), and so the chips had to undergo significant processing to
build up these vias into ?nanoelectrodes?. This required several deposition, etching,
114
and polishing steps [116]. In this work and in [43], no post-processing of the CMOS
ICs was necessary, making device integration much simpler and cheaper.
8.2.5 Temporal Resolution and Hardware Interface
The ability to make real-time capacitance measurements throughout multi-day
biological experiments in the cell culture environment is critical in order to make a
case for using CMOS sensors in lieu of conventional biological assays. Most of the
state of the art work presented in Fig. 8.2 meet the sub-minute temporal resolution
needed to track cell morphological changes. However in the work that presented
experiments where cytotoxic effects of drugs were examined, the measurements are
relatively sparse (panel (c), [11]). This may be attributed to the supervised manner
in which data had to be collected.
While most of the state of the art capacitance sensor designs operate using
the CS, CBCM, or CTF mechanisms, the hardware interfaces needed for chip read-
out vary quite widely. The chips may require equipment such as data acquisition
cards, FPGAs, source-meters, and digital oscilloscopes. These types of hardware
can be cumbersome to transport and maintain in non-laboratory settings, making
these systems susceptible to increased noise and interference in clinical settings and
ill-suited for applications requiring small form factor devices with simple readout
interfaces. In contrast, the design presented here requires only a microcontroller,
making it suitable for applications outside the traditional laboratory setting. Table
8.1 presents a summary of existing work.?
??CSensed: the peak change in capacitance after addition of cells DAC - Digital to analog
115
converter, DAQ - Data acquisition, DDS - Direct digital synthesizer, DLA - Digital logic analyzer,
DO - Digital oscilloscope, FPGA - Field programmable gate array, PC - Personal computer, SM -
Source-Meter, ?C - Microcontroller
116
[18, 19, 9, 43] [11, 12] [10] [116, 13] [48] This work.
Method CS CBCM CS CBCM CTF CTF
Tech. (?m) 0.5 0.35 0.25 0.09 0.25 0.35
Array Size 28 8? 8 16? 16 256? 256 1 4? 4,4? 8
Biospecies MDA-MB-231 HEK293 kidney, S. epidermidis BEAS, THP1, S. epidermidis A2780, CP70
breast cancer H1299 lung cancer MCF7 cancer ovarian cancer
Readout DAQ Card FPGA + PC DO + DLA + PC FPGA + DDS ?C, DO ?C
System + PC + SM + DC Supply + DAC
Samp. Rate 1 kHz 1 - 100 kHz 37 frames/s 4.5 frames/s - 0.18 - 45 Hz
ADC Off-chip Sigma-Delta Off-chip Cyclic Off-chip Counter
Sensitivity - 320 mV/fF 55 mV/fF - - 590 kHz/fF
(mV/fF)
Res. (af) 135 10 450 1 - 16
Electrode 202, 302, 402 502 14? 16 ?0.092 1002, 2002 302
Size (?m2)
?CSensed? 8,12,25 fF 6.6 fF 310 aF (5 cells at 50 aF 53, 445 aF 1 fF
(200 kCells/ml) 62 aF/cell)
?CSensed/Area 13 - 20 2.63 1.38 1960 5.3, 11.1 1.10, 1.28
(af/?m2)
Table 8.1: Biosensors from State of Art
8.2.6 Imaging on CMOS ICs
Imaging cells on chip concurrently with data recording is an important method
of validating sensor measurements. Both fluorescence-based [85, 86, 84] and reflec-
tive [11, 37] microscopy have been used in prior work to image cells on top of
CMOS chips. However the imaging was done at discrete times using standard mi-
croscopy and did not achieve the temporal resolution to correlate with real-time cell
morphology changes. Furthermore, imaging was performed outside the incubator,
which meant measurements had to be interrupted and the cells had to be removed
from the controlled culture environment.
In this work the packaged devices were modified to include two wells, one for
cell culture and imaging, and the other to allow gas exchange and maintenance of
pH levels in media. The in-incubator assembly allowed for high-resolution imaging
with the ability to discriminate individual cells in different phases of growth and
death. To our knowledge, this is the first time that a CMOS-based sensor has been
used and validated concurrently with time-lapse images of the sensor surface.
8.3 Future Work
8.3.1 Increased Spatial Density
A research track that can be extended from this dissertation is an increase
in the spatial density of sensors, i.e. a high-density capacitance sensor array. As
discussed in Section 3.6, the generation 1 chip consisted of a 4 ? 4 array of pixels
118
arranged with an X and Y pitch of 196 ?m and 186 ?m, respectively, and covering
an area of an area of 618 ? 588 ?m2. The electrodes themselves are 30 ? 30 ?m2 in
size and so the actual sensing area covered by the electrodes is roughly 120 ? 120
?m2, meaning a fill factor of 4.0 %.
As an immediate goal, the area occupied by each sensor could first be mini-
mized. The presented capacitance sensor consists of the input inter-digitated elec-
trodes and requisite CTF circuitry. Shielding is required to protect the circuitry
from ambient light and so one layer of metal is typically dedicated to creating a
light shield. In the fabricated chips (which used a 3 metal CMOS process) this was
done with M3. That means the area of each sensor is now expanded to include
both the electrodes and circuitry. If a process node with more metal layers is avail-
able, however, the circuitry itself could be situated underneath the electrodes, with
shielding created using a lower metal layer. This would allow bringing electrodes
from multiple sensors as close together as possible. One limitation to this would be
the fringing effects that would occur from the top-metal layers between sensors that
are nearby.
An arrangement of sensors with a high spatial density will allow the creation
of larger arrays. An important consideration now would be the readout architecture.
Each sensor in the array is scanned sequentially, with a finite amount of time required
for each sampling, leading to a linear increase in overall sampling time as the number
of sensors increases. The generation 2 and 3 chips developed in this work aimed to
address this by including parallel readout of sensors, and random-access selection
of sensors, respectively. Additionally, the programmable integration time could be
119
used to trade off resolution for sampling speed.
8.3.2 Microfluidic Integration
As in many LoC systems, microlfuidic integration is a natural extension to
the platform that could expand the capabilities of the sensor for use in scientifically
relevant experiments. For instance, microfluidic channels could be fabricated on
top of the sensing array to generate a linear or logarithmic gradient of dilutions
of chemotherapeutic agents [117, 118]. The capacitance sensor could then be used
to explore cytotoxic effects to cells in a high-throughput manner. The drug ad-
ministration profile could also be given a temporal characteristic (e.g. pulsing or
ramping) to investigate effects on cells that would not be possible with traditional
laboratory-based cytotoxicity assays. In addition, multiple microfluidic channels
could be incorporated to expose cells to a multi-drug dose regimen.
8.3.3 Closed-loop Feedback
Another extension related to microfluidic integration would be to incorporate
feedback controls to perform automated drug dispensing, as seen in Fig. 8.3. For
instance the capacitance measurements could be automatically analyzed to track
proliferation, and once a threshold level is reached, a cytotoxic agent could be dis-
pensed into the culture area through a perfusion system. Furthermore, drug dis-
pensing could be regulated by incorporating cell cycle marker detection in order to
perform automated cell cycle-targeted drug screening therapies.
120
Perfusion 
System
Yes
Micro-
Sensor Detect
controller
No
Figure 8.3: Block diagram for a feedback control loop to perform automated drug dis-
pensement.
8.3.4 Sensitivity Improvement
As discussed in Section 3.1, the sensed capacitance at the input electrodes is
the series combination of the passivation layer capacitance (CPSV ) and the input
capacitance of interest (CIN), as given by the following equation:
C ? 3CPSVCINsense = C0 + (8.1)
CPSV + 2CIN
where C0 is the sum of the parasitics giving rise to the baseline capacitance level.
Therefore, if CPSV is much larger than CIN , then its contribution to the sensed
capacitance will be diminished. In the ideal case where CPSV  CIN , then the
second term of Equation 8.1 reduces to 3CIN .
One way to increase CPSV is to thin the passivation layer. This could be
performed using standard wet or dry etching of the layer, although care must be
121
taken to stop before the metal electrodes are exposed. To analyze this effect, an
FEM model was built and simulated using the structure presented in Section 3.2.
Fig. 8.4 shows FEM simulation results of sensed capacitance as a function of cell
radius. Panels (a) shows the change in capacitance for a passivation thickness of
1.6 ?m, and panel (b) shows the result when the thickness is reduced to 0.8 ?m.
As can be seen, there is a larger change in capacitance as cell-substrate coupling
increases with a thinner passivation layer (?500 aF vs. ?1000 aF). A linear fit of
the data points results in a slope of -23 aF/?m and -59 aF/?m for the two cases,
respectively.
Another way to increase CPSV is to increase the dielectric constant of the
passivation layer. This can be achieved by first opening holes over the electrodes,
either through post-processing or through the CMOS foundry glass cut option, and
then depositing a thin film of a high-K dielectric such as hafnium oxide.
a) b)
25.50 31.80
25.30 31.60
25.10 31.40
24.90 31.20
24.70 31.00
24.50 30.80
24.30 30.60
0 5 10 15 20 25 0 5 10 15 20 25
Cell radius (?m) 
Figure 8.4: FEM simulation results showing the effect of passivation layer thinning on
sensed capacitance. The electrode and cell structure presented in Section 3.2 was simulated
as cell radius was increased with a passivation thickness of a) 1.6 ?m, and (b) 0.8 ?m.
122
Capacitance (fF)
8.3.5 Surface Functionalization
The presented CMOS sensor was designed to monitor cell-phenomena through
the dielectric changes that they produce at the cell-substrate interface. However, this
sensing mechanism can be generalized using labels to detect other biomarkers in an
unknown sample solution. For instance, the sensor surface could be functionalized
with molecules that are specific towards a target biomarker. When the surface is
then exposed to the sample to be interrogated, biomarker molecules in the sample are
captured and immobilized close to the sensor electrodes. This could then modulate
the dielectric properties at the sensor surface, which can then be detected as a change
in capacitance. Therefore, this type of lab-on-CMOS sensor could potentially be
used for detection of affinity-based binding of other biomarkers, not just for cell-
based assays.
8.4 Applications
8.4.1 Drug Development - Electronic Multi-well Plate
Cell-based assays are the standard diagnostic tools used in the drug devel-
opment process. In contrast to biochemical assays, cell-based assays offer a more
complex model for study that can generate more biologically-relevant information
to predict the response of an organism to a particular substance.
Researchers currently perform cell-based assays in multi-well plates that can
include 6, 12, 24, 48, or 96 culture wells on one platform to perform multiple ex-
123
periments in parallel. The type of screening assay depends on a number of fac-
tors including the cell type being used, the parameter being measured, chemistry
limitations, and the exposure/measurement times involved. However, the process
generally works as follows: 1) seed cells into the well plate, 2) add the test agent
after a certain growth period, 3) allow any agent-induced effects to occur (several
hours to days), 4) add a reagent or marker, 5) measure the response.
The last step is typically done using a multi-well plate reader that can operate
using one of several detection methods, including colorimetric, fluorometric, and
bioluminsecent mechanisms. However, as discussed in Section 2.2.2, these detection
methods are end-point assays. Parallel experiments are needed to obtain a temporal
sequence of discrete sample points, and temporal accuracy would be limited by the
variability across samples and the sampling technique. Much richer information
about the dynamics of cell growth would be available from a real-time measurement
methodology.
Therefore the sensor presented in this work could be used as an alternative
to traditional multi-well assays. Fig. 8.5 shows a diagram of an ?electronic multi-
well plate? that can be created using the lab-on-CMOS capacitance sensor chip.
Starting with a standard-sized well, a hole can be laser-cut into the bottom of the
well to create an opening that is sized to the chip?s sensing array area. The chip is
then attached to the underside of the well and sealed with a passivation material
such as a biocompatible silicone or epoxy. The electronic connections to the chip
are made underneath the plate and chips across a single row are all connected
together. Multiple rows can then be connected together a single bus line that goes
124
Multi-well plate
Well Opening
CMOS Sensor Daughterboard Connector to
microcontroller
Figure 8.5: Schematic of an ?electronic multi-well plate?. Each well contains a chip with
its sensing area exposed to the cell culture. Electronic connections are made underneath
the well plate across columns and rows on a daughterboard. The final routing is wired to
a connector that can then be read using a microcontroller.
to a connector port on a daughterboard. This is where the microcontroller readout
board is connected.
With this system, researchers can perform the same high-throughput parallel
experiments currently performed but glean much more temporally-intensive data
due to the real-time measurement capabilities of the capacitance sensor. Further-
more, this obviates the existing need to perform parallel experiments to obtain the
temporal profile of cellular responses. A single experiment with the chip/well could
replace the multiple experiments required quantify at different time points; for in-
stance if response data is required every hour, this is a 24? reduction over a single
day.
125
8.4.2 Point-of-care Diagnostics and Personalized Medicine
As described previously, the capacitance sensor has potential to be used for
detection of affinity-based binding of other biomarkers. This can be explored by per-
forming surface functionalization of the sensor electrodes to target specific analytes.
This could then have potential use in point-of-care diagnostic applications, such as
at doctor?s offices where samples from patients can be analyzed in a high-throughput
manner.
The presented cell monitoring sensor could further be used in the emerging
field of personalized medicine, where different treatment programs are developed for
different patients in order to maximize drug efficacy while minimizing other toxic
effects. For instance, tumor cells from a patient can be harvested and cultured on
the device, which could then be implanted back in to the patient?s body. The chip
could then provide real-time viability information about the cancer cells within the
patient?s body, while they undergo chemotherapeutic treatment. The treatment
could then be adjusted for the individual to increase drug efficacy and safety, and
enhance patient outcomes.
8.5 Publication Record
8.5.1 Refereed Journal Articles
1. B. Senevirathna, S. Lu, M. P. Dandin, J. Basile, E. Smela, and P. A. Ab-
shire, ?Real-Time Measurements of Cell Proliferation Using a Lab-on-CMOS
Capacitance Sensor Array,? IEEE Trans. Biomed. Circuits Syst., vol. 12, no.
3, pp. 510-520, 2018.
126
2. B. Senevirathna, S. Lu, M. P. Dandin, E. Smela, and P. A. Abshire, ?Corre-
lation of Capacitance and Microscopy Measurements Using Image Processing
for a Lab-on-CMOS Microsystem,? IEEE Trans. Biomed. Circuits Syst., 2019.
[Accepted]
3. B. Senevirathna, S. Lu, M. P. Dandin, J. Basile, E. Smela, and P. A. Abshire,
?High resolution monitoring of chemotherapeutic agent potency in cancer cells
using a CMOS capacitance biosensor,? Biosensors and Bioelectronics, 2019.
[Accepted]
4. J. Kilpija?rvi, N. Halonen, M. Sobocinski, A. Hassinen, B. Senevirathna, K.
Uvdal, P. Abshire, E. Smela, S. Kellokumpu, J. Juuti and A. Lloyd Spetz,
?LTCC Packaged Ring Oscillator Based Sensor for Evaluation of Cell Prolif-
eration,? Sensors, vol. 18, no. 10, p. 3346, 2018.
8.5.2 Journal Articles in Preparation
1. B. Senevirathna, S. Lu, M. P. Dandin, J. Basile, E. Smela, and P. A. Abshire,
?Quantifying mitotic activity for cancer prognosis using a CMOS biosensor,?
Lab Chip, 2019. [In preparation]
8.5.3 Refereed Conference Proceedings
1. B. Senevirathna, S. Lu, N. Renegar, M. Dandin, E. Smela, and P. Abshire,
?System on a Chip for Automated Cell Assays using a Lab-on-CMOS Plat-
form,? IEEE Int. Symp. Circ. Syst., 2019.
2. S. Lu, B. Senevirathna, M. Dandin, E. Smela, and P. Abshire, ?System
Integration of IC chips for Lab-on-CMOS Applications,? IEEE Int. Symp.
Circ. Syst., 2018.
3. B. Senevirathna, S. Lu, and P. Abshire, ?Characterization of a High Dy-
namic Range Lab-on-CMOS Capacitance Sensor Array,? IEEE Int. Symp.
Circ. Syst., 2017.
4. B. Senevirathna, A. Castro, M. Dandin, E. Smela, and P. Abshire, ?Lab-on-
CMOS Capacitance Sensor Array for Real-Time Cell Viability Measurements
with I2C Readout,? IEEE Int. Symp. Circ. Syst., 2016.
5. B. Senevirathna, A. Castro, T. Datta-Chaudhuri, E. Smela, and P. Abshire,
?Characterization of an Active Micro-Electrode Array with Spike Detection
and Asynchronous Readout,? IEEE Int. Midwest Symp. Circ. Syst., 2017.
127
6. T. Datta-Chaudhuri, B. Senevirathna, A. Castro, E. Smela, and P. Ab-
shire, ?An active micro-electrode array with spike detection and asynchronous
readout,? IEEE Biomed. Circ. Syst. Conf., 2014.
7. B. Senevirathna, and P. Abshire, ?Spatio-temporal compressed sensing for
real-time wireless EEG monitoring,? IEEE Int. Symp. Circ. Syst., 2018.
8. B. Senevirathna, and P. Abshire, ?Power Analysis of a Mobile EEG System
with Compressed Sensing,? IEEE Biomed. Circ. Syst. Conf., 2017.
9. B. Senevirathna, L. Berman, N. Bertoni, F. Pareschi, M. Mangia, R. Rovatti,
G. Setti, J. Simon, and P. Abshire, ?Low cost mobile EEG for characterization
of cortical auditory responses,? IEEE Int. Symp. Circ. Syst., 2016.
10. N. Bertoni, B. Senevirathna, F. Pareschi, M. Mangia, R. Rovatti, P. Abshire,
J. Z. Simon, and G. Setti, ?Low-power EEG monitor based on compressed
sensing with compressed domain noise rejection,? IEEE Int. Symp. Circ.
Syst., 2016.
128
Appendix A
Appendix
A.1 Sensor Design Details
The analog voltage used to bias the oscillators is created using an on-chip
Beta-multiplier reference circuit (Fig. A.1, [119]). Transistors M1-4 and R form
the base Beta-multiplier circuit. M5-8 form a simple amplifier to provide feedback
to increase the output resistance of M4. M9-14 form a cascoded current source to
generate the required bias level. MS1-S3 form a start-up circuit.
The on-chip control logic was designed to sequentially cycle through all chan-
nels in the capacitance sensor array. Fig. A.2 shows a diagram of the finite-state
machine implemented on chip generation 3. When the reset signal is pulsed, the chip
MP M
M 2 VO
M 1 M14S1 M9 M11
M5 M6
M MS3 10 M15
M12
MS2 M3 M7 M8 M4 M16
M13
R
Figure A.1: Circuit schematic of on-chip bias generator.
129
goes to the IDLE state and INIT states that initialize the internal logic registers.
The integration time and active pixel registers are programmed through I2C. The
logic then begins cycling through each pixel in the array. In the the CHECK PX
state where the current pixel is checked to see whether it is active, if not the pixel
is skipped (SKIP state).
If the pixel is active, the logic goes into the SCAN state where the oscillator
is enabled through row/column decoders and the counter is activated. After the
integration time is reached, the logic goes into COLLECT which flags the DRDY
signal and loads the counter value into the I2C data registers. An additional CLEAR
state is used to allow for data latching and resetting the active pixel tracker. The
chip then goes into the TX WAIT state where it sits idle until an I2C read cycle is
performed, after which it moves to the next pixel.
IDLE
reset
ALL 
STATES
INIT
!activeChans[px_addr]
px_addr++
!doneAccumulating
SKIP CHECK_PX drdy = 1
SCAN
ack_received activeChans[px_addr]
doneAccumulating
TX_WAIT
COLLECT
drdy = 0
!ack_received CLEAR
px_addr = (px_addr == 31) ? 0 : px_addr++  
Figure A.2: Finite-state machine diagram of on-chip logic.
130
Fig. A.3 shows simulation results of the readout timing for a single pixel (the
timing has been scaled down for simulation purposes). Each sensor is activated and
allowed to start up for a fixed length of time (200 ns), after which the counter is
allowed to increment. After the integration time (1 ?s), the counter data is latched
into the I2C data registers, and the data ready signal is flagged to trigger a read
through the microcontroller.
System CLK
Reset
Pixel Output
Counter
I2C Data In
Data Ready
(a)
System CLK
Reset
Pixel Output
Counter
I2C Data In
Data Ready
(b)
Figure A.3: Simulation results of logic for readout of one pixel (accumulation time short-
ened to 1 ?s for simulation purposes. The figure depicts a a) full view, and b) zoomed-in
view.
131
A.2 Sample Data
This section presents a few additional examples of data from in vitro ex-
periments. Contact Bathiya Senevirathna (bsenevir@umd.edu) or Pamela Abshire
(pabshrie@umd.edu) for full dataset, or visit our website (isr.umd.edu/IBIS).
1000
1
2
500
3
4
0
1000
5
6
500
7
8
0
1000
9
10
500
11
12
0
1000
13
14
500
15
16
0
0 10 20 30 40 50 60 70
Time (hours)
(a)
(b) left (c) right
Figure A.4: a) Response curves of the chip as CP70 cells adhere and proliferate across the
sensor array. The first two vertical black lines at t = 25 hours and t = 47 hours indicate
times when data recording was paused for imaging. Right panel shows image of surface
of chip after b) 25 hours, and c) 47 hours.
132
| C| (aF)
1000
1
2
500
3
4
0
1000
5
6
500
7
8
0
1000
9
10
500
11
12
0
1000
13
14
500
15
16
0
0 10 20 30 40 50 60 70 80
Time (hours)
(a)
(b) left (c) right
Figure A.5: a) Response curves of the chip as A2780 cells adhere and proliferate across
the sensor array. The first two vertical black lines at t = 28 hours and t = 49.5 hours
indicate times when data recording was paused for imaging. Right panel shows image of
surface of chip after b) 28 hours, and c) 49.5 hours.
133
| C| (aF)
1000
1
2
500
3
4
0
1000
5
6
500
7
8
0
1000
9
10
500
11
12
0
1000
13
14
500
15
16
0
0 10 20 30 40 50 60 70 80
Time (hours)
(a)
(b) left (c) right
Figure A.6: a) Response curves of the chip as CP70 cells adhere and proliferate across the
sensor array. The first two vertical black lines at t = 28 hours and t = 48 hours indicate
times when data recording was paused for imaging. Right panel shows image of surface
of chip after b) 28 hours, and c) 48 hours.
134
| C| (aF)
 
a) b) 
 
c) 
 
 
Figure A.7: Results of experiment on Device 1 with A2780 cells. Microscope image of
chip surface after a) 21.5 hours and b) 46 hours of incubation. c) Response curves where
sensors with and without cell coverage are grouped together. Two sensors observed to
have no coverage at the two time intervals (2,4) and (4,3) are shown in dotted red, with
their mean as a solid line. The other sensors are shown in dotted black, with mean as a
solid line.
135
 
 
 
 a) b) 
 
c) 
 
 
Figure A.8: Results of experiment with A27 80 cells. Microscope image of chip surface
after a) 27 hours and b) 70 hours of incubat ion. c) Response curves of sensors with cell
 
coverage in gray, with the mean as a solid black line.
136
Bibliography
[1] Jeong-Woo Lee, Dong-Jin Min, Jiyoun Kim, and Wonchan Kim. A 600-dpi ca-
pacitive fingerprint sensor chip and image-synthesis technique. IEEE Journal
of Solid-State Circuits, 34(4):469?475, 1999.
[2] J C Chen, D Sylvester, and Chenming Hu. An on-chip, interconnect capaci-
tance characterization method with sub-femto-farad resolution. IEEE Trans-
actions on Semiconductor Manufacturing, 11(2):204?210, may 1998.
[3] Michael S.C. Lu, Yi Chung Chen, and Po Chiun Huang. 5x5 CMOS capacitive
sensor array for detection of the neurotransmitter dopamine. Biosensors and
Bioelectronics, 26(3):1093?1097, 2010.
[4] Harvey Lodish, Arnold Berk, S. Lawrence Zipursky, Paul Matsudaira, David
Baltimore, and James E. Darnell. Molecular Cell Biology. W.H. Freeman,
New York, 4 edition, 2000.
[5] I Giaever and C R Keese. Micromotion of mammalian cells measured electri-
cally. Proceedings of the National Academy of Sciences of the United States of
America, 88(17):7896?7900, 1991.
[6] Timir Datta-Chaudhuri, Elisabeth Smela, and Pamela A. Abshire. System-
on-Chip Considerations for Heterogeneous Integration of CMOS and Flu-
idic Bio-Interfaces. IEEE Transactions on Biomedical Circuits and Systems,
10(6):1129?1142, 2016.
[7] Humayun Irshad, Antoine Veillard, Ludovic Roux, and Daniel Racoceanu.
Methods for nuclei detection, segmentation, and classification in digital
histopathology: A review-current status and future potential. IEEE Reviews
in Biomedical Engineering, 7:97?114, 2014.
[8] Thermo Scientific. alamarBlue Cell Viability Assay Reagent. Technical report,
2012.
[9] Somashekar Bangalore Prakash and Pamela Abshire. Tracking cancer cell pro-
liferation on a CMOS capacitance sensor chip. Biosensors and Bioelectronics,
23(10):1449?1457, 2008.
[10] Numa Couniot, Laurent A Francis, and Denis Flandre. A 16x16 CMOS Ca-
pacitive Biosensor Array Towards Detection of Single Bacterial Cell. IEEE
Transactions on Biomedical Circuits and Systems, 10(2):364?374, 2016.
[11] Ghazal Nabovati, Ebrahim Ghafar-Zadeh, Antoine Letourneau, and Mohamad
Sawan. Smart Cell Culture Monitoring and Drug Test Platform Using CMOS
Capacitive Sensor Array. IEEE Transactions on Biomedical Engineering,
66(4), 2019.
137
[12] Ghazal Nabovati, Ebrahim Ghafar-zadeh, Antoine Letourneau, and Mohamad
Sawan. Towards High Throughput Cell Growth Screening: A New CMOS 8
x 8 Biosensor Array for Life Science Applications. IEEE Transactions on
Biomedical Circuits and Systems, 11(2):380?391, 2017.
[13] C. Laborde, F. Pittino, H. A. Verhoeven, S. G. Lemay, L. Selmi, M. A.
Jongsma, and F. P. Widdershoven. Real-time imaging of microparticles
and living cells with CMOS nanocapacitor arrays. Nature Nanotechnology,
10(9):791?795, 2015.
[14] Sigma-Aldrich. Cell Viability and Proliferation Assays, 2019.
[15] D. L. Coutu and T. Schroeder. Probing cellular processes by long-term live
imaging - historic problems and current solutions. Journal of Cell Science,
126(17):3805?3815, 2013.
[16] Yehya H. Ghallab and Yehea Ismail. CMOS based lab-on-a-chip: Applications,
challenges and future trends. IEEE Circuits and Systems Magazine, 14(2):27?
47, 2014.
[17] Ka Meng Lei, Pui In Mak, Man Kay Law, and Rui P. Martins. CMOS biosen-
sors for: In vitro diagnosis-transducing mechanisms and applications. Lab on
a Chip, 16(19):3664?3681, 2016.
[18] Somashekar Bangalore Prakash and Pamela Abshire. A CMOS capacitance
sensor that monitors cell viability. IEEE Sensors, 2005:1177?1180, 2005.
[19] Somashekar Bangalore Prakash, Pamela Abshire, Mario Urdaneta, and Elis-
abeth Smela. A CMOS capacitance sensor for cell adhesion characterization.
In IEEE International Symposium on Circuits and Systems, pages 3495?3498,
2005.
[20] Elmore Susan. Apoptosis: A Reveiw of Programmed Cell Death. Toxicologic
Pathology, 35(4):496?516, 2007.
[21] D Sylvester, J C Chen, and Chenming Hu. Investigation of interconnect capac-
itance characterization using charge-based capacitance measurement (CBCM)
technique and three-dimensional simulation. IEEE Journal of Solid-State Cir-
cuits, 33(3):449?453, mar 1998.
[22] M Tartagni and R Guerrieri. A fingerprint sensor based on the feedback
capacitive sensing scheme. IEEE Journal of Solid-State Circuits, 33(1):133?
142, jan 1998.
[23] Somashekar Bangalore Prakash and Pamela Abshire. A fully differential rail-
to-rail CMOS capacitance sensor with floating-gate trimming for mismatch
compensation. IEEE Transactions on Circuits and Systems I: Regular Papers,
56(5):975?986, 2009.
138
[24] Ebrahim Ghafar-Zadeh and Mohamad Sawan. Charge-based capacitive sensor
array for CMOS-based laboratory-on-chip applications. IEEE Sensors Jour-
nal, 8(3-4):325?332, 2008.
[25] Ghazal Nabovati, Ebrahim Ghafar-Zadeh, Maryam Mirzaei, Giancarlo Ayala-
Charca, Falah Awwad, and Mohamad Sawan. A New Fully Differential CMOS
Capacitance to Digital Converter for Lab-on-Chip Applications. IEEE Trans-
actions on Biomedical Circuits and Systems, 9(3):353?361, 2014.
[26] Amelia Ahmad Khalili and Mohd Ridzuan Ahmad. A Review of cell adhesion
studies for biomedical and biological applications. International Journal of
Molecular Sciences, 16(8):18149?18184, 2015.
[27] Micah Dembo and Yu-Li Wang. Stresses at the Cell-to-Substrate Interface dur-
ing Locomotion of Fibroblasts. Biophysical Journal, 76(4):2307?2316, 1999.
[28] V K Gupta, Ihab A Sraj, Konstantinos Konstantopoulos, and Charles D Eggle-
ton. Multi-scale simulation of L-selectin?PSGL-1-dependent homotypic leuko-
cyte binding and rupture. Biomechanics and Modeling in Mechanobiology,
9(5):613?627, oct 2010.
[29] Kevin V Christ, Kyle B Williamson, Kristyn S Masters, and Kevin T
Turner. Measurement of single-cell adhesion strength using a microfluidic
assay. Biomedical Microdevices, 12(3):443?455, jun 2010.
[30] A Rollan, D McCormack, L McHale, H McCormack, and A P McHale. A rapid
in situ, colorimetric assay for the determination of mammalian cell viability
in alginate-immobilized and encapsulated systems. Bioprocess Engineering,
15(1):47?49, jun 1996.
[31] Anja Schreer, Cheryl Tinson, James P Sherry, and Kristin Schirmer. Appli-
cation of Alamar blue/5-carboxyfluorescein diacetate acetoxymethyl ester as
a noninvasive cell viability assay in primary hepatocytes from rainbow trout.
Analytical Biochemistry, 344(1):76?85, 2005.
[32] R A Hoebe, C H Van Oven, T W J Gadella, P B Dhonukshe, C J F Van
Noorden, and E M M Manders. Controlled light-exposure microscopy re-
duces photobleaching and phototoxicity in fluorescence live-cell imaging. Nat
Biotech, 25(2):249?253, feb 2007.
[33] Lavanya Balasubramanian, Kay-Pong Yip, Tai-Hsin Hsu, and Chun-Min Lo.
Impedance analysis of renal vascular smooth muscle cells. American journal
of physiology. Cell physiology, 295:C954?C965, 2008.
[34] J Wegener, a Janshoff, and H J Galla. Cell adhesion monitoring using a quartz
crystal microbalance: comparative analysis of different mammalian cell lines.
European biophysics journal : EBJ, 28(1):26?37, 1999.
139
[35] Michael Saitakis and Electra Gizeli. Acoustic sensors as a biophysical tool for
probing cell attachment and cell/surface interactions. Cellular and Molecular
Life Sciences, 69(3):357?371, 2012.
[36] M.L. Pourciel-Gouzy, W. Sant, I. Humenyuk, L. Malaquin, X. Dollat, and
P. Temple-Boyer. Development of pH-ISFET sensors for the detection of bac-
terial activity. Sensors and Actuators B: Chemical, 103(1-2):247?251, 2004.
[37] Kiichi Niitsu, Shoko Ota, Kohei Gamo, Hiroki Kondo, Masaru Hori, and
Kazuo Nakazato. Development of Microelectrode Arrays Using Electroless
Plating for CMOS-Based Direct Counting of Bacterial and HeLa Cells. IEEE
Transactions on Biomedical Circuits and Systems, 9(5):607?619, 2015.
[38] Ivar Giaever and Charles Richard Keese. Monitoring fibroblast behavior in tis-
sue culture with an applied electric field. Proceedings of the National Academy
of Sciences of the United States of America, 81(12):3761?3764, 1984.
[39] I Giaever and C R Keese. A morphological biosensor for mammalian cells.
Nature, 366(6455):591?592, 1993.
[40] Ivar Giaever and Charles R Keese. Use of Electric Fields to Monitor the
Dynamical Aspect of Cell Behavior in Tissue Culture. IEEE Transactions on
Biomedical Engineering, BME-33(2):242?247, 1986.
[41] Daniel Opp, Brian Wafula, Jennifer Lim, Eric Huang, Jun-Chih Lo, and Chun-
Min Lo. Use of electric cellsubstrate impedance sensing to assess in vitro
cytotoxicity. Biosensors and Bioelectronics, 24(8):2625?2629, 2009.
[42] Diane S Allen-Gipson, Matthew C Zimmerman, Hui Zhang, Glenda Castel-
lanos, Jennifer K O?Malley, Horacio Alvarez-Ramirez, Kusum Kharbanda,
Joseph H Sisson, and Todd A Wyatt. Smoke Extract Impairs Adenosine
Wound Healing. Implications of Smoke-Generated Reactive Oxygen Species.
American Journal of Respiratory Cell and Molecular Biology, 48(5):665?673,
2013.
[43] S. B. Prakash and P. Abshire. On-chip capacitance sensing for cell monitoring
applications. IEEE Sensors Journal, 7(3):440?447, 2007.
[44] Ebrahim Ghafar-Zadeh, Mohamad Sawan, Vamsy P. Chodavarapu, and
Tahereh Hosseini-Nia. Bacteria growth monitoring through a differential
CMOS capacitive sensor. IEEE Transactions on Biomedical Circuits and Sys-
tems, 4(4):232?238, 2010.
[45] Somashekar Bangalore Prakash and Pamela Abshire. A fully differential rail-
to-rail capacitance measurement circuit for integrated cell sensing. Proceedings
of IEEE Sensors, pages 1444?1447, 2007.
140
[46] Somashekar Bangalore Prakash and Pamela Abshire. A fully differential
CMOS capacitance sensor design, testing and array architecture. Proceedings
- IEEE International Symposium on Circuits and Systems, pages 165?168,
2008.
[47] Timir Datta, Emily Naviasky, and Pamela Abshire. Floating-gate Capacitance
Sensor Array for Cell Viability Monitoring. IEEE Biomedical Circuits and
Systems Conference (BioCAS), pages 65?68, 2013.
[48] Numa Couniot, David Bol, Olivier Poncelet, Laurent A Francis, and Denis
Flandre. A Capacitance-to-Frequency Converter With On-Chip Passivated
Microelectrodes for Bacteria Detection in Saline Buffers Up to 575 MHz. IEEE
Transactions on Circuits and Systems II: Express Briefs, 62(2):159?163, 2015.
[49] Alessia Damilano, Marco Crepaldi, Paolo Motto Ros, and Danilo Demarchi.
A 130 nm event-driven voltage and temperature insensitive capacitive ROC.
In 7th Euromicro Conference on Digital System Design, DSD 2014, pages
663?666, 2014.
[50] Timir Datta-Chaudhuri, Pamela Abshire, and Elisabeth Smela. Packag-
ing commercial CMOS chips for lab on a chip integration. Lab on a chip,
14(10):1753?66, 2014.
[51] Emily Naviasky, Timir Datta-Chaudhuri, and Pamela Abshire. High resolu-
tion capacitance sensor array for real-time monitoring of cell viability. IEEE
International Symposium on Circuits and Systems, pages 634?637, 2014.
[52] Inc Applied BioPhysics. ECIS - Applied Biophysics, 2017.
[53] V. Tsouti, C. Boutopoulos, I. Zergioti, and S. Chatzandroulis. Capacitive
microsystems for biological sensing. Biosensors and Bioelectronics, 27(1):1?
11, 2011.
[54] N. Couniot, A. Afzalian, N. Van Overstraeten-Schlo?gel, L. A. Francis, and
D. Flandre. Capacitive biosensing of bacterial cells: Analytical model and
numerical simulations. Sensors and Actuators, B: Chemical, 211:428?438,
2015.
[55] Bathiya Senevirathna, Alexander Castro, Marc Dandin, Elisabeth Smela, and
Pamela Abshire. Lab-on-CMOS Capacitance Sensor Array for Real- Time Cell
Viability Measurements with I2C Readout. IEEE International Symposium
on Circuits and Systems, pages 2863?2866, 2016.
[56] Bathiya Senevirathna, Sheung Lu, and Pamela Abshire. Characterization
of a High Dynamic Range Lab-on-CMOS Capacitance Sensor Array. IEEE
International Symposium on Circuits and Systems (ISCAS), pages 1?4, 2017.
141
[57] Bathiya Prashan Senevirathna, Sheung Lu, Marc P. Dandin, John Basile,
Elisabeth Smela, and Pamela A. Abshire. Real-Time Measurements of Cell
Proliferation Using a Lab-on-CMOS Capacitance Sensor Array. IEEE Trans-
actions on Biomedical Circuits and Systems, 12(3):510?520, 2018.
[58] Boris Andreev, Edward L Titlebaum, and EBY G Friedman. Sizing CMOS
Inverters with Miller Effect and Threshold Voltage Variations. Journal of
Circuits, Systems, and Computers, World Scientific, 15(3):437?454, 2006.
[59] Asad A. Abidi. Phase noise and jitter in CMOS ring oscillators. IEEE Journal
of Solid-State Circuits, 41(8):1803?1816, 2006.
[60] Yu Shiang Lin, Dennis Sylvester, and David Blaauw. An ultra low power 1V,
220nW temperature sensor for passive wireless applications. In Proceedings of
the Custom Integrated Circuits Conference, pages 507?510, 2008.
[61] Steve Fielding. I2C Slave - Minimalist I2C slave IP core that provides the
basic framework for the implementation of custom I2C slave devices, 2013.
[62] University of Washington. Dielectric constants of common solvents.
[63] Tsung Hsueh Lee and Pamela A. Abshire. Frequency-Boost Jitter Reduction
for Voltage-Controlled Ring Oscillators. IEEE Transactions on Very Large
Scale Integration (VLSI) Systems, 24(10):3156?3168, 2016.
[64] Thermo Scientific. Thermo Scientific Forma Series 3 Water Jacketed CO2
Incubator, 2015.
[65] Microchip. MCP1703 Datasheet, 2010.
[66] Sheung Lu, Bathiya Senevirathna, Marc Dandin, Elisabeth Smela, and Pamela
Abshire. System Integration of IC chips for Lab-on-CMOS Applications. 2018
IEEE International Symposium on Circuits and Systems (ISCAS), pages 2?6,
2018.
[67] Niina Halonen, Joni Kilpija?rvi, Maciej Sobocinski, Timir Datta-Chaudhuri,
Antti Hassinen, Someshekar B. Prakash, Peter Mo?ller, Pamela Abshire, Sakari
Kellokumpu, and Anita Lloyd Spetz. Low temperature co-fired ceramic pack-
aging of CMOS capacitive sensor chip towards cell viability monitoring. Beil-
stein Journal of Nanotechnology, 7(1):1871?1877, 2016.
[68] Joni Kilpija?rvi, Niina Halonen, Maciej Sobocinski, Antti Hassinen, Bathiya
Senevirathna, Kajsa Uvdal, Pamela Abshire, Elisabeth Smela, Sakari Kel-
lokumpu, Jari Juuti, and Anita Lloyd Spetz. LTCC Packaged Ring Oscillator
Based Sensor for Evaluation of Cell Proliferation. Sensors, 18(10):3346, 2018.
[69] Ryan J Petrie, Andrew D Doyle, and Kenneth M Yamada. Random versus
directionally persistent cell migration. Nature reviews. Molecular cell biology,
10(8):538?549, 2009.
142
[70] Richard Kasprowicz, Rakesh Suman, and Peter O?Toole. Characterising live
cell behaviour: Traditional label-free and quantitative phase imaging ap-
proaches. International Journal of Biochemistry and Cell Biology, 84:89?95,
2017.
[71] Fuyong Xing and Lin Yang. Robust nucleus/cell detection and segmentation
in digital pathology and microscopy images: A comprehensive review. IEEE
Reviews in Biomedical Engineering, 9:234?263, 2016.
[72] Leica Microsystems. Immersion Objectives: Using Oil, Glycerol, or Water to
Overcome some of the Limits of Resolution, 2017.
[73] Rachel Bareither and David Pollard. A review of advanced small-scale parallel
bioreactor technology for accelerated process development: Current state and
future need. Biotechnology Progress, 27(1):2?14, 2011.
[74] Di Jin, Dennis Wong, Junxiang Li, Zhang Luo, Yiran Guo, Bifeng Liu, Qiong
Wu, Chih-Ming Ho, and Peng Fei. Compact Wireless Microscope for In-
Situ Time Course Study of Large Scale Cell Dynamics within an Incubator.
Scientific Reports, 5(1):18483, 2016.
[75] Nikon Instruments Inc. Nikon BioStation IM-Q, 2017.
[76] World Precision Instruments. Microscope Environmental Chamber, Heat and
CO2 controller, 2017.
[77] GE Healthcare Life Sciences. Cytell Cell Imaging System, 2017.
[78] Leica Microsystems. Inverted Microscope Platform Leica DMi8, 2017.
[79] Marco Ballini, Jan Muller, Paolo Livi, Yihui Chen, Urs Frey, Alexander Stet-
tler, Amir Shadmani, Vijay Viswam, Ian Lloyd Jones, David Jackel, Milos
Radivojevic, Marta K. Lewandowska, Wei Gong, Michele Fiscella, Douglas J.
Bakkum, Flavio Heer, and Andreas Hierlemann. A 1024-channel CMOS mi-
croelectrode array with 26,400 electrodes for recording and stimulation of elec-
trogenic cells in vitro. IEEE Journal of Solid-State Circuits, 49(11):2705?2719,
2014.
[80] Jelena Dragas, Vijay Viswam, Amir Shadmani, Yihui Chen, Raziyeh Bounik,
Alexander Stettler, Milos Radivojevic, Sydney Geissler, Marie Engelene J.
Obien, Jan Mu?ller, and Andreas Hierlemann. In Vitro Multi-Functional Mi-
croelectrode Array Featuring 59 760 Electrodes, 2048 Electrophysiology Chan-
nels, Stimulation, Impedance Measurement, and Neurotransmitter Detection
Channels. IEEE Journal of Solid-State Circuits, 52(6):1576?1590, 2017.
[81] A. Lambacher, M. Jenkner, M. Merz, B. Eversmann, R. A. Kaul, F. Hofmann,
R. Thewes, and P. Fromherz. Electrical imaging of neuronal activity by multi-
transistor-array (MTA) recording at 7.8 ?m resolution. Applied Physics A:
Materials Science and Processing, 79(7):1607?1611, 2004.
143
[82] K. Imfeld, A. Garenne, S. Neukom, A. Maccione, S. Martinoia, M. Koudelka-
Hep, and L. Berdondini. High-resolution MEA platform for in-vitro electro-
genic cell networks imaging. IEEE Engineering in Medicine and Biology -
Proceedings, pages 6085?6088, 2007.
[83] Jong Seok Park, Moez Karim Aziz, Sensen Li, Taiyun Chi, Sandra Ivonne
Grijalva, Jung Hoon Sung, Hee Cheol Cho, and Hua Wang. 1024-Pixel CMOS
Multimodality Joint Cellular Sensor/Stimulator Array for Real-Time Holistic
Cellular Characterization and Cell-Based Drug Screening. IEEE Transactions
on Biomedical Circuits and Systems, 12(1):80?94, 2018.
[84] Taiyun Chi, Jong Seok Park, Jessica Butts, Tracy Hookway, Amy Su, Chengjie
Zhu, Mark Styczynski, Todd McDevitt, and Hua Wang. A multimodality
CMOS sensor array for cell-based assay and drug screening. IEEE Transac-
tions on Biomedical Circuits and Systems, 9(6):801?814, 2015.
[85] Jong Seok Park, Sandra I Grijalva, Moez K Aziz, Taiyun Chi, Sensen Li,
Michael N Sayegh, Adam Wang, Hee Cheol Cho, and Hua Wang. Multi-
parametric cell profiling with a CMOS quad-modality cellular interfacing array
for label-free fully automated drug screening. Lab on a Chip, 18(19):3037?
3050, 2018.
[86] Carolina Mora Lopez, Ho Sung Chun, Shiwei Wang, Laurent Berti, Jan
Putzeys, Carl Van Den Bulcke, Jan Willem Weijers, Andrea Firrincieli, Veerle
Reumers, Dries Braeken, and Nick Van Helleputte. A multimodal CMOS
MEA for high-throughput intracellular action potential measurements and
impedance spectroscopy in drug-screening applications. IEEE Journal of
Solid-State Circuits, 53(11):3076?3086, 2018.
[87] Jeffrey Abbott, Tianyang Ye, Ling Qin, Marsela Jorgolli, Rona S. Gertner,
Donhee Ham, and Hongkun Park. CMOS nanoelectrode array for all-electrical
intracellular electrophysiological imaging. Nature Nanotechnology, 12(5):460?
466, 2017.
[88] Erik Meijering. Cell Segmentation: 50 Years Down the Road. IEEE Signal
Processing Magazine, 29(5):140?145, 2012.
[89] Nobuyuki Otsu. A Threshold Selection Method from Gray-Level Histograms.
IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62?66, 1979.
[90] D Andreou, P Reichardt, M Werner, M Schuler, B Jobke, D Pink, G Gosheger,
F Traub, and P U Tunn. Prognostic relevance of the mitotic count and the
amount of viable tumour after neoadjuvant chemotherapy for primary, lo-
calised, high-grade soft tissue sarcoma. British Journal of Cancer, 112(3):455?
460, 2014.
144
[91] Mitko Veta, Paul J. Van Diest, Mehdi Jiwa, Shaimaa Al-Janabi, and
Josien P.W. Pluim. Mitosis counting in breast cancer: Object-level inter-
observer agreement and comparison to an automatic method. PLoS ONE,
11(8):1?8, 2016.
[92] Michael Bonert and Angela J. Tate. Mitotic counts in breast cancer should
be standardized with a uniform sample area. BioMedical Engineering Online,
16(1):1?8, 2017.
[93] Jean-michel Coindre. Grading of soft tissue sarcomas. Archives of pathology
& laboratory medicine, 130(10):1448?1453, 2006.
[94] Humayun Irshad, Sepehr Jalali, Ludovic Roux, Daniel Racoceanu, Lim Joo
Hwee, Gilles Le Naour, and Fre?de?rique Capron. Automated mitosis detec-
tion using texture, SIFT features and HMAX biologically inspired approach.
Journal of Pathology Informatics, 4(2):12, 2013.
[95] Seungil Huh, Dai Fei Elmer Ker, Ryoma Bise, Mei Chen, and Takeo Kanade.
Automated Mitosis Detection of Stem Cell Populations in Phase-Contrast
Microscopy Images. IEEE Transactions on Medical Imaging, 30(3):586?596,
2010.
[96] Antoine Veillard, Maria S Kulikova, and Daniel Racoceanu. Cell nuclei ex-
traction from breast cancer histopathologyimages using colour, texture, scale
and shape information. Diagnostic Pathology, 8(S1):8?10, 2013.
[97] Christopher Malon, Elena Brachtel, Eric Cosatto, Hans Peter Graf, Atsushi
Kurata, Masahiko Kuroda, John S. Meyer, Akira Saito, Shulin Wu, and
Yukako Yagi. Mitotic figure recognition: Agreement among pathologists and
computerized detector. Analytical Cellular Pathology, 35(2):97?100, 2012.
[98] Ashkan Tashk, Mohammad Sadegh Helfroush, Habibollah Danyali, and Mo-
jgan Akbarzadeh. An automatic mitosis detection method for breast cancer
histopathology slide images based on objective and pixel-wise textural features
classification. Conference on Information and Knowledge Technology (ITK),
pages 406?410, 2013.
[99] Dan C Cirean, Alessandro Giusti, Luca M Gambardella, and Ju?rgen Schmid-
huber. Mitosis Detection in Breast Cancer Histology Images with Deep Neural
Networks BT - Medical Image Computing and Computer-Assisted Interven-
tion MICCAI 2013. Proceedings of Medical Image Computing and Computer-
Assisted Intervention (MICCAI), pages 411?418, 2013.
[100] Monjoy Saha, Chandan Chakraborty, and Daniel Racoceanu. Efficient deep
learning model for mitosis detection using breast histopathology images. Com-
puterized Medical Imaging and Graphics, 64(March 2017):29?40, 2018.
145
[101] Alessandro Giusti, Claudio Caccia, Dan C. Ciresari, Jurgen Schmidhuber, and
Luca M. Gambardella. A comparison of algorithms and humans for mitosis
detection. 2014 IEEE 11th International Symposium on Biomedical Imaging
(ISBI), pages 1360?1363, 2014.
[102] David O. Morgan. The Cell Cycle: Principles of Control. New Science Press,
London, 2007.
[103] M.A. Dickson and G.K. Schwartz. Development of cell-cycle inhibitors for
cancer therapy. Current Onocology, 16(2):36?43, 2009.
[104] Carmen Dominguez-Brauer, Kelsie L. Thu, Jacqueline M. Mason, Heiko
Blaser, Mark R. Bray, and Tak W. Mak. Targeting Mitosis in Cancer: Emerg-
ing Strategies. Molecular Cell, 60(4):524?536, 2015.
[105] Charles J. Sherr and Jiri Bartek. Cell CycleTargeted Cancer Therapies. An-
nual Review of Cancer Biology, 1(1):41?57, 2016.
[106] Perry J Kaufman. Trading Systems and Methods. John Wiley & Sons, Incor-
porated, Somerset, UNITED STATES, 2013.
[107] Christina L. Dix, Helen K. Matthews, Marina Uroz, Susannah McLaren, Lu-
cie Wolf, Nicholas Heatley, Zaw Win, Pedro Almada, Ricardo Henriques,
Michael Boutros, Xavier Trepat, and Buzz Baum. The Role of Mitotic Cell-
Substrate Adhesion Re-modeling in Animal Cell Division. Developmental Cell,
45(1):132?145, 2018.
[108] John G. Lock, Matthew C. Jones, Janet A. Askari, Xiaowei Gong, Anna
Oddone, Helene Olofsson, Sara Go?ransson, Melike Lakadamyali, Martin J.
Humphries, and Staffan Stro?mblad. Reticular adhesions are a distinct class
of cell-matrix adhesions that mediate attachment during mitosis. Nature Cell
Biology, 20(11):1290?1302, 2018.
[109] Manuela F. Azzola, Helen M. Shaw, John F. Thompson, Seng-jaw Soong,
Richard A. Scolyer, Geoffrey F. Watson, Marjorie H. Colman, and Yuting
Zhang. Tumor mitotic rate is a more powerful prognostic indicator than ulcer-
ation in patients with primary cutaneous melanoma. Cancer, 97(6):1488?1498,
2003.
[110] Benoite Me?ry, Jean-Baptiste Guy, Alexis Vallard, Sophie Espenel, Dominique
Ardail, Claire Rodriguez-Lafrasse, Chloe? Rancoule, and Nicolas Magne?. In
Vitro Cell Death Determination for Drug Discovery: A Landscape Review of
Real Issues. Journal of Cell Death, 10, 2017.
[111] E III. Frei and JP Eder. Combination Chemotherapy. BC Decker, Hamilton,
6th edition, 2003.
[112] Robert A Gatenby, Ariosto S Silva, Robert J Gillies, and B Roy Frieden.
Adaptive Therapy. Cancer Research, 69(11):4894?4903, 2009.
146
[113] B. Zhao, M. T. Hemann, and D. A. Lauffenburger. Intratumor heterogene-
ity alters most effective drugs in designed combinations. Proceedings of the
National Academy of Sciences, 111(29):10773?10778, 2014.
[114] John O?Brien, Ian Wilson, Terry Orton, and Franc?ois Pognan. Investigation of
the Alamar Blue (resazurin) fluorescent dye for the assessment of mammalian
cell cytotoxicity. European Journal of Biochemistry, 267(17):5421?5426, 2000.
[115] Thermo Scientific. alamarBlue Assay for Cell Viability, 2019.
[116] Frans Widdershoven, Andrea Cossettini, Cecilia Laborde, Andrea Bandiziol,
Paul P Van Swinderen, Serge G Lemay, and Luca Selmi. A CMOS Pixe-
lated Nanocapacitor Biosensor Platform for High-Frequency Impedance Spec-
troscopy and Imaging. IEEE Transactions on Biomedical Circuits and Sys-
tems, 12(6):1369?1382, 2018.
[117] Glenn M. Walker, Nancy Monteiro-Riviere, Jillian Rouse, and Adrian T.
O?Neill. A linear dilution microfluidic device for cytotoxicity assays. Lab
on a Chip, 7(2):226?232, 2007.
[118] Choong Kim, Kangsun Lee, Jong Hyun Kim, Kyeong Sik Shin, Kyu Jung Lee,
Tae Song Kim, and Ji Yoon Kang. A serial dilution microfluidic device using
a ladder network generating logarithmic or linear concentrations. Lab on a
Chip, 8(3):473?479, 2008.
[119] R Baker. CMOS : Circuit Design, Layout, and Simulation. IEEE Press/Wiley,
Hoboken, N.J, third edition, 2010.
147