Lab on a Chip

View Article Online

COMMUNICATION View Journal | View Issue

Assessing red blood cell deformability from

Cite this: Lab Chip, 2022, 22, 26
microscopy images using deep learning†
Received 6th November 2021, Erik S. Lamoureux,ab Emel Islamzada,bc Matthew V. J. Wiens, d
Accepted 27th November 2021 Kerryn Matthews, ab Simon P. Duffyabe and Hongshen Ma *abdf
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

DOI: 10.1039/d1lc01006a

rsc.li/loc

Red blood cells (RBCs) must be highly deformable to transit through Introduction
the microvasculature to deliver oxygen to tissues. The loss of RBC
deformability resulting from pathology, natural aging, or storage in Red blood cells (RBCs) are highly specialized cells that
blood bags can impede the proper function of these cells. A variety facilitate tissue respiration by delivering oxygen and removing
of methods have been developed to measure RBC deformability, but carbon dioxide.1,2 RBCs transverse through the entire
these methods require specialized equipment, long measurement circulatory system approximately every 60 seconds. Their
time, and highly skilled personnel. To address this challenge, we journey includes the microvasculature, where RBCs must
investigated whether a machine learning approach could be used to deform through capillaries measuring as little as 2 μm in
predict donor RBC deformability based on morphological features diameter, as well as the inter-endothelial clefts of the spleen
from single cell microscope images. We used the microfluidic measuring 0.5–1.0 μm in diameter.3,4 The loss of RBC
ratchet device to sort RBCs based on deformability. Sorted cells are deformability, due to pathology, natural aging, or storage in
then imaged and used to train a deep learning model to classify RBC blood bags, reduces the ability of RBCs to circulate and
based image features related to cell deformability. This model facilitate their removal from circulation by phagocytes in the
correctly predicted deformability of individual RBCs with 81 ± 11% spleen and the liver.5,6 As a result, there is significant interest
accuracy averaged across ten donors. Using this model to score the in methods for measuring RBC deformability as a potential
deformability of RBC samples was accurate to within 10.4 ± 6.8% of biomarker for diseases, such as malaria7 and
2,8
the value obtained using the microfluidic ratchet device. While hemoglobinopathies, or for assessing the quality of donated
machine learning methods are frequently developed to automate RBCs for use in blood transfusions.9,10
human image analysis, our study is remarkable in showing that deep Approaches for measuring RBC deformability can be
learning of single cell microscopy images could be used to assess classified as either flow-based or deformation-based
RBC deformability, a property not normally measurable by imaging. methods. Flow-based methods deform RBCs using fluid
Measuring RBC deformability by imaging is also desirable because it shear stress and then measure the resulting shape change. A
can be performed rapidly using a standard microscopy system, classical method is ektacytometry, which deforms RBCs using
potentially enabling RBC deformability studies to be performed as shear flow between two transparent cylinders and then uses
part of routine clinical assessments. optical diffraction to measure the resulting population RBC
elongation.11,12 Other flow-based methods deform RBCs
using high shear flow through microchannels and then
a
Department of Mechanical Engineering, University of British Columbia, 2054- measure the resulting RBC elongation using high speed
6250 Applied Science Lane, Vancouver, BC, V6T 1Z4, Canada. imaging13,14 or electrical impedance.15 Classical deformation-
E-mail: [email protected]
b
based methods include micropipette aspiration,16 atomic
Centre for Blood Research, University of British Columbia, Vancouver, BC,
Canada
force microscopy,17 and optical tweezers,18,19 which measure
c
Department of Pathology and Laboratory Medicine, University of British RBC deformability via single cell manipulation, and require
Columbia, Vancouver, BC, Canada complex experimentation, skilled personnel, and specialized
d
School of Biomedical Engineering, University of British Columbia, Vancouver, BC, equipment.9 Microfluidic deformation-based methods
Canada
e
measure RBC deformability via capillary obstruction,20
British Columbia Institute of Technology, Burnaby, BC, Canada
f
Vancouver Prostate Centre, Vancouver General Hospital, Vancouver, BC, Canada
deposition length in tapered constrictions,21,22 transit
† Electronic supplementary information (ESI) available. See DOI: 10.1039/ pressure through constrictions,7,10,23–27 transit time through
d1lc01006a constrictions,28–30 and sorting RBCs based on deformability

26 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

using microfluidic ratchets.31–33 A common challenge for all deformability. We show that the deep learning model can
existing deformability assays is the needs for specialized classify RBCs into deformable or rigid fractions using donor
apparatus and skilled personnel, which limit the ability to dataset sizes ranging from 20 000 to 70 000 images. For a
translate the technology to clinical settings.34 Additionally, sample of ten donors, who were diverse in terms of blood
since different assays rely on different underlying principles type and sex, testing classification accuracy ranged from 64–
for measuring RBC deformability, it is often difficult or 95% with an aggregate mean (± SD) of 81 ± 11%. Using our
impossible to compare results across studies. model to predict donor RBC rigidity scores (RS) was accurate
As an alternative to physical measurement of RBC to within a mean of 10.4 ± 6.8% deviation, compared to
deformability, cues for biophysical changes in these cells measurement using a microfluidic device. Our results
may be obtainable from microscopy images, without the need confirm that RBC deformability can be assessed from
for highly specialized equipment and personnel. RBCs microscopy images to potentially simplify the assessment of
typically exhibit a highly deformable biconcave discoid RBC deformability.
morphology, and deviation from this morphology may
correspond with changes in cell deformability.35,36 In fact,
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

deep learning methods have been developed to assess Results

changes in RBC morphology during cold storage,37 Approach
malaria,38–43 sickle cell disease,44–49 and thalassemia.50–52
However, RBC morphology varies over the life cycle of the cell Our experimental approach (Fig. 1) involves first using the
and this variability may obscure efforts to infer deformability microfluidic ratchet device to sort RBCs based on
from cell morphology. Furthermore, no specific deformability in order to acquire training data for a deep
morphological features can be directly attributed to learning image classification model. The microfluidic ratchet
predictable changes in RBC deformability. We recently device (Fig. 1A) sorts RBCs based on deformability by
developed a microfluidic process for deformability-based squeezing cells through a matrix of tapered constrictions
sorting of RBCs31–33 as well as a deep learning method to (described in Fig. 2). After microfluidic sorting (Fig. 1A),
distinguish cell lines based on feature differences RBCs from each outlet are extracted from the microfluidic
imperceptible to human cognition.53 We hypothesized that a device and placed in wells on a 96-well plate (Fig. 1B). These
combination of these advances could enable the indirect wells are then imaged using an optical microscope in
measurement of RBC deformability via cell imaging from brightfield using a 40× objective (Fig. 1C and D). The
optical microscopy to identify image-based cell microscopy images are processed and segmented into single
morphological features associated with deformability. cell images (Fig. 1E). The resulting datasets are then used to
Here, we investigate the potential to use deep learning to train and test a convolutional neural network for
assess RBC deformability based on cell morphological deformability-based image classification (Fig. 1F). Finally, the
features from brightfield microscopy images. We leverage the trained deep learning model is used to classify RBCs in a test
ability of a microfluidic ratchet device to sort RBCs based on set based on deformability. The aggregate RBC deformability
deformability to generate training sets of RBCs with distinct is also used to obtain a deep learning-derived deformability

Fig. 1 Overall experimental approach. (A) Deformability based sorting using the microfluidic ratchet device. (B) Sorted RBC fractions are
transferred to a well plate for imaging. (C) Brightfield imaging using a 40× objective. (D) Example full well image scan. (E) Examples of individual
segmented RBCs from donor 1. (F) Structure of the convolutional neural network for image-based cell classification. (G) RBC rigidity score
estimated using the CNN. (H) Rigidity score measured by deformability-based cell sorting.

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 27
 View Article Online

Communication Lab on a Chip

Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

Fig. 2 Microfluidic ratchet sorting device operation. (A) Tapered funnel constrictions allow for unidirectional filtration of RBCs upwards through
the constrictions based on their tapered geometry and upward oscillatory flow. Downward oscillatory flow declogs cells that are too rigid to pass
through the constrictions. The constriction width for a row of tapers decreases along the flow path (i.e., a < b) to enable deformability-based cell
separation. (B) Deformability based sorting occurs in a matrix of tapered constrictions, producing a ratcheting effect. (C) 10× magnification
micrograph of the active sorting region where cells are routed to different deformability outlets. (D) 4× magnification micrograph of the entire
microfluidic sorting region. RBCs are pictured flowing in an upward diagonal direction from the sample flow inlet on the left, through the matrix of
constrictions, and are primarily routed to outlets 3 and 4 on the right.

profile of the RBC sample (Fig. 1G), which is compared to the RBC rigidity score (RS)
profile obtained by microfluidic testing (Fig. 1H). After deformability-based sorting, the RBC distribution can
be shown as a histogram, where a rightward distribution
Deformability based cell sorting corresponds to a more rigid RBC sample (Fig. 3A). To
compare deformability between samples, the RBC
We sorted RBCs based on deformability using the distribution can be shown as a cumulative distribution,
microfluidic ratchet device described and validated which allows us to define a rigidity score (RS) as the outlet
previously.7,9,10,16,31–33,54–57 Briefly, RBCs migrate under number where the cumulative distribution crosses 50%.
oscillatory flow through a matrix of micropores with Fractional outlet numbers can be determined by linear
openings ranging from 1.5 μm to 7.5 μm (Table 1). The interpolation between data points greater and less than 50%
micropores have the same opening in each row, but in the cumulative distribution function. RBC samples from
progressively smaller openings along each column different donors showed significant variability in their RS
(Fig. 2A and B). When the cells can no longer transit the value. For example, from the ten donors in this study the RS
micropores along a particular row, they instead flow along ranged from 2.47 to 3.50 (Fig. 3B and C).
the row and are directed into one of 12 outlets
(Fig. 2B and C). Oscillatory flow within the device dislodges
cells that are blocked by the micropores, ensuring that the Donor-based variability in RBC deformability
cells do not clog the device, allowing for unimpeded Blood donations from ten healthy donors were obtained and
operation. In this way, the RBC sample is fractionated based RBC samples were sorted based on deformability using the
on deformability at a rate of ∼600 cells per minute. More microfluidic ratchet device (Fig. 3C). The donors were diverse
detailed device design and operational details can be found in terms of blood type and sex (Table 3). Three blood samples
in our previous work.9,55,58 were fresh from a citrate tube (used the day of donation),

Table 1 Constriction size for each outlet number

Outlet 1 2 3 4 5 6 7 8 9 10 11 12
Size (μm) 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 3.50 4.50 5.50 7.50

28 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

Fig. 3 Microfluidic deformability-based sorting results. (A) RBC distribution after deformability-based sorting for select donors. Donor 2 is the
most deformable sample (orange) and donor 4 is the most rigid sample (green) of the ten donors analyzed. (B) Cumulative distribution of RBC
deformability from donor 2 and 4. The rigidity scores (RS) are measured at the 50% crossover of the cumulative distribution. (C) Cumulative
distributions and RS for all ten donors.

four were fresh from a blood bag (≤3 days after donation), 2.36–3.69 for fresh blood and blood stored for one week
one was in the second week of blood bag storage, one was in could reach elevated RS scores of 3.74.9
the second week of tube storage, and one was stored in a
blood bag for just over 3 weeks (Table 3). Donor RBCs were Optical microscopy imaging for deep learning
sorted into outlets 1–6, with the majority (>99%) sorted to
After deformability-based cell sorting, the sorted cells are
outlets 2–5. The cumulative distribution of sorted cells to
extracted from the microfluidic device by pipetting and
each outlet is presented in Fig. 3C. These deformability
placed in 96-well imaging plate (Fig. 1B). Samples from each
curves and RS are donor-specific and can be reliably
outlet were split in half and placed in two wells to introduce
measured in repeated experiments using replicate
additional variance in the imaging conditions. These
microfluidic devices.9 The donor RS range from 2.47–3.50,
variations include a greater variety of light conditions based
which is similar to previous results where the RS ranged from
on location of cells in the well, cells with different
thicknesses of suspension fluid due to its meniscus, and
Table 2 Number of unique segmented single cell images in each outlet different imaging conditions resulting from differences in
for all ten donors. From these datasets, deformable (outlets 2 and 3) and focus and exposure time. Full image scans were conducted
rigid (outlets 4 and 5) images are split for training and testing, and are on each well using a 40× objective and a DS-Qi2 camera on a
subsequently augmented for class balancing
Nikon Ti-2E inverted microscope, capturing brightfield
Number of images images of 2424 × 2424 pixels (Fig. 1C and D). Image captures
Donor Outlet 2 Outlet 3 Outlet 4 Outlet 5 near the edge of the wells were often out of focus and were
1 — 14 655 13 831 —
discarded prior to segmentation.
2 7566 14 099 7195 —
3 7849 15 352 14 152 3448 Segmentation
4 — 5345 3634 13 482
5 — 10 583 9879 — To perform deep learning classification of individual RBCs,
6 7931 26 514 26 407 10 606 we developed a Python program to extract 60 × 60 pixel image
7 — 8717 25 455 5938
patches each containing a single RBC (Fig. 1E). Cells are
8 9845 18 124 16 776 6948
9 — 17 094 13 231 — located using a Sobel operator for edge detection and Otsu
10 — 24 276 26 811 18 789 multi-thresholding. A centre of mass measurement is

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 29
 View Article Online

Communication Lab on a Chip

Table 3 Donor characteristics, deep learning results, and comparison of microfluidic and deep learning determined rigidity scores (RS)

Blood Storage type, storage Training Validation Testing Microfluidic Deep

Donor type Sex time Outlets accuracyb (%) accuracyb (%) accuracyc (%) RS learning RS
1 A− F Fresh, day 0 3, 4 92 93 92 3.27 3.19
2 A+ M Tube, day 12 2, 3, 4 96 95 95 2.47 2.66
3 O− M Bag, day 1 2, 3, 4, 5 86 83 84 2.96 2.84
4 A+ M Bag, day 2 3, 4, 5 99 96 83 3.50 3.41
5 B+ F Bag, day 2 3, 4 86 84 82 3.15 3.18
6 B+ M Bag, day 3 2, 3, 4, 5 76 67 64 2.96 3.05
7 B+ M Fresh, day 0 3, 4, 5 85 75 71 3.47 3.24
8 B+ F Bag, day 9 2, 3, 4, 5 80 68 66 2.77 2.75
9 Oa F Fresh, day 0 3, 4 96 93 95 3.27 3.22
10 B+ F Bag, day 23 3, 4, 5 87 83 81 3.25 3.11
a
Donor Rhesus factor unknown. b Training and validation accuracies indicate the average accuracy across the five cross validation folds.
c
Additional testing metrics (including precision, recall, F1-score, and ROC AUC) are tabulated in Table S1.†
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

conducted to centre the identified cell for image cropping. adjusted based on training time and training and validation
After the cells are identified and cropped, they proceed convergence outcomes.
through a selection algorithm that keeps images with a single
cell centered in the image and rejects images with multiple Training and validation
cells. Further, the resulting selected cropped images are
The CNN was trained using single-cell images with true
manually audited and any remaining images with multiple
deformability labels determined by microfluidic
cells or those out of focus are removed. This segmentation
deformability-based cell sorting. The CNN utilized balanced
procedure resulted in datasets containing 20 000 to 70 000
training classes of 10 000 images per class per donor. Cell
single cell images for each donor (Table 2). The segmentation
images were augmented by a random integer multiple of 90-
process was a significant bottleneck in our analysis process,
degree rotation to capture different cell orientations and
with the manual data cleaning procedure requiring multiple
lighting characteristics (Fig. 1E). Classes with fewer than
hours of manual auditing per donor. This bottleneck could
10 000 images were up-sampled, and classes with greater
potentially be addressed in future iterations of this approach
than 10 000 images were sub-sampled. The model was
by implementing a U-Net algorithm for pixel-based image
trained and validated using five-fold cross validation. The
segmentation.59
average training accuracies across the five folds for each
donor is shown in Table 3 and Fig. 4D. Using the validation
Network design and its convergence for each fold, the model was evaluated
for hyperparameter tuning iteratively to determine the
We designed a convolutional neural network (CNN) to
optimal learning rate, number of epochs, optimizer type, and
conduct image feature extraction and classification using the
batch size. Learning rates and number of epochs were donor-
Keras library in TensorFlow (Fig. 1F). For feature extraction,
specific and ranged from 0.0001 to 0.1 and 25 to 80,
the model utilizes a series of 4 convolutional layers and 3
respectively. We settled on a stochastic gradient descent
max pooling layers. We tried a variety of initial kernel sizes
(SGD) operator with decay 10−6 and Nesterov momentum 0.9,
as larger sizes capture more robust image features. We
as an Adam optimizer did not markedly improve outcomes,
settled on an initial convolutional layer kernel size of 7 × 7 as
and the SGD hyperparameters were reliable across all donors.
larger sizes did not substantially improve performance. The
A batch-size of 32 was determined and held for all donors
second convolutional layer kernel size is 5 × 5, and the final
because larger batches can result in reduced ability to
two are 3 × 3. Each convolutional layer is followed by batch
generalize, and smaller batches can make learning too
normalization and ReLU activation. The latter classification
stochastic, causing unreliable convergence. There was
section consists of 3 fully connected layers and a final
significant variation in the model's ability to converge during
smaller fully connected output layer. The 3 fully connected
training between different donors, illustrated by the variation
layers are followed by batch normalization, ReLU activation,
in training epochs and learning rates.
and 20% dropout. The output layer uses a SoftMax error
function for backpropagation during training. The network
utilizes a binary cross-entropy loss function and stochastic Classification
gradient descent for optimization. The design of this model We initially used our CNN to classify cells from microscopy
was influenced by the AlexNet model architecture60 and deep images based on the outlet they were sorted to. However,
learning architecture used in previous work in our lab.53 The classifying cells in this manner resulted in poor classification
model was modified for 60 × 60 pixel input images and the accuracies (Fig. 4A). This result likely derives from there
number of layers and their sizes were initially iteratively being substantially fewer single cell images from outlets 2

30 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

Fig. 4 (A) Normalized confusion matrix for donor 3 when RBCs from outlets 2–5 were considered as separate classes. (B) Normalized confusion
matrix for donor 3 when RBCs from outlets 2 and 3 (deformable) were pooled into a single class, and RBCs from outlets 4 and 5 (rigid) were
pooled into a single class. Classification accuracy is greatly increased with outlets pooled. (C) Example images from outlets 2–5 for donor 6. There
does not seem to be obvious visual differences in the RBCs from different outlets. (D) Image classification training, validation, and testing
accuracies for all donors. Training and validation accuracies are averaged over the five folds.

and 5, compared to outlets 3 and 4 (Table 2). To create large and robust dataset is required for training and testing
balanced classes of 10 000 images, substantial up-sampling (Fig. 4C). Using this classification scheme, training
was performed on cells from outlets 2 and 5, requiring many accuracies ranged 76–99% and final validation accuracies
repeated cells. As a result, the variety of cell images seen by ranged 67–96%, shown in Table 3 and Fig. 4D.
the model for these classes were significantly limited.
Interestingly, more misclassification occurred between
adjacent outlets (Fig. 4A), indicating that the model was Testing
learning some common deformability-based cell features. The testing datasets are comprised of 20% of the overall
To improve our classification accuracy, we collapsed the segmented data and were separated from the training and
image data from outlets 2 and 3 together and outlets 4 and 5 validation sets prior to augmentation. This process ensured
together to create classes of deformable and rigid cells. By that there are no cell image repeats between the training and
collapsing the classes in this manner, the datasets are more testing sets. For each donor, the images were augmented
robust as additional cell images are available for using the same method used in training and were up-
augmentation, which required less up-sampling. This binary sampled or down-sampled to obtain balanced testing sets of
classification method resulted in substantially improved 2000 images per class. Among our donor population, we
deformability image predictions (Fig. 4B). An additional observed testing accuracies ranging 64–95% with an
advantage of this binary classification method is that inter- aggregate mean (± SD) of 81 ± 11% (Table 3 and Fig. 4D). For
donor comparisons are more appropriate as all donors have each donor, we observed that testing accuracy corresponded
cells sorted to outlets 3 and 4, but not all have cells sorted to well with validation accuracy. Additional testing metrics,
outlets 2 or 5. Since differences in deformability is not a including precision, recall, F1-score, and ROC AUC can be
feature that can be easily discerned by a human observer, a found in Table S1.†

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 31
 View Article Online

Communication Lab on a Chip

Saliency maps We further investigated how classification accuracy varies

A key consideration for image classification using deep with reduced training data. Classification accuracy were
learning is whether classification was driven by imaging found to decrease predictably with reduced training data
artifacts, such as lighting, sample preparation, image (Fig. S4†). Specifically, our results indicate that ideally >5000
acquisition parameters, and position in the imaging well.61 images per class are required to obtain robust classification
To resolve this potential issue, we generated saliency maps to accuracy.
assess if the model learned relevant cell morphological Finally, we performed a Hough circle transform analysis
features. A saliency map is a visual representation of the on cells from different deformability outlets to investigate
spatial support of a particular class to indicate which pixels whether deformability could be determined from planar cell
in a given image had the greatest effect on the classification size. We found planar cell size were invariant to cell
probability.62 As shown in Fig. 5, the saliency map shows that deformability as determined by sorting, but interestingly,
pixels having the most influence on classification there were statistically significant differences in cell size
corresponded to those clustered around the cell itself, between donors, which confirmed the validity of our Hough
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

especially distinct cell features, rather than surrounding circle transform analysis (Fig. S2†).
regions. This result confirms that our model is classifying
RBC deformability based on cell morphological features. Using deep learning to determine rigidity scores of RBC
samples

Model performance analysis By classifying RBCs into deformable and rigid fractions, we
can use this result to estimate the RS for each RBC sample.
To further confirm the model performs image classification RBCs classified as deformable and rigid classes are assigned
using cellular features and is not strongly influenced by to outlet 3 and 4, respectively. This scheme is justified for the
artifacts, we investigated how classification accuracy varies ten donors studied here since the vast majority (86%) of cells
with reduced imaging resolution. Images were down-sampled from all these donors were sorted into outlets 3 (39%) and 4
from 60 × 60 pixel images to mimic imaging at lower (47%), with the reminder sorted into outlets 2 (5%) and 5
resolutions. Classification accuracy were found to reduce (9%). This approach also ensured the RS calculations are
predictably with increased down-sampling (Fig. S3†). consistent for all donors as not every donor had cells sorted

Fig. 5 Original images (left), saliency maps (centre), and saliency maps with smoothing (right) of randomly sampled RBCs from donor 4. The
saliency map indicates the strength of pixel contributions to the final classification output by the CNN. Warmer colours indicate greater
contribution and cooler colours indicate lesser contribution.

32 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

to outlets 2 or 5, but all donors had cells sorted to outlets 3 deviated between a minimum 0.02 (donor 8) to a maximum
and 4. After classification, the RS is then calculated as before 0.23 (donor 7) with a mean of 0.10 ± 0.07. Using the potential
by linearly interpolating the cumulative distribution deep learning RS range between 2.50 and 3.50, the resulting
deformability curve to find the outlet number at the 50% mean percent deviation between the deep learning estimated
crossover frequency. Assigning cells from outlets 2 and 3 to RS and the microfluidic RS across the ten donors is 10.4 ±
outlet 3, outlets 4 and 5 to outlet 4 can potentially 6.8%. Previous work with this microfluidic device has shown
overestimate the RS for deformable samples and a standard deviation for RS of 0.17 across five different tests
underestimate the RS for rigid samples, but this error is on the same sample.9 Expressed differently, this RS deviates
small and systematic since the majority of the RBCs from by 13.8%, relative to the RS range across all donors in that
healthy donors are sorted into outlets 3 and 4. work, indicating the level of deviation seen here is within
Comparing the cumulative distributions and RS obtained acceptable variation in RS resulting from random sampling
by cell sorting using the microfluidic ratchet device with the and manufacturing. Furthermore, we plotted the RS acquired
RS estimated by deep learn showed a strong agreement by deep learning against RS acquired by microfluidics for the
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

(Fig. 6). Specifically, the measured and estimated RS values ten donors and found a high degree of correlation between

Fig. 6 (A–J) Comparison of microfluidic (solid lines) and deep learning (dashed lines) derived RBC deformability cumulative distributions and RS
for all ten donors. (K) Relationship between rigidity scores determined by microfluidics and deep learning methods for all ten donors. The deep
learning RS values are strongly correlated to the microfluidic RS values (r = 0.94) and this relationship is statistically significant (p < 0.0001).

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 33
 View Article Online

Communication Lab on a Chip

the two methods (Fig. 6K), with a Pearson's correlation of r = able to use deep learning to classify RBCs based on
0.94 and p < 0.0001. deformability with a mean testing accuracy of 81%.
To assess whether cell sorting using the microfluidic Furthermore, we used this approach to estimate the rigidity
ratchet device may have altered the RBCs, we used the donor- scores (RS) of a RBC sample, which deviated by a mean of
specific trained CNN to classify unsorted cells from donor 2, 10.4% from physical measurement using the microfluidic
3, and 4. This test also assessed the model's generalizability ratchet device. We previously showed significant inter-donor
by applying a testing dataset acquired in a separate variability in RBC deformability,9 which were successfully
experimental procedure from the training data. Donors 2, 3, captured by our deep learning model. To ensure our
and 4 were selected for assessment as these donors represent inference of RBC deformability was robust to factors like well
the full range of donor RS: donor 2 is most deformable (RS = location, lighting conditions, and presence of debris, we
2.47), donor 3 is in the middle (RS = 2.96) and donor 4 is the introduced additional data variance by purposefully dividing
most rigid (RS = 3.50). As before, the cumulative distribution RBC specimens into multiple imaging wells and augmenting
and RS obtained by cell sorting and deep learning were cell images by rotation during database creation. The
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

similar with the difference in RS for donor 2, 3, and 4 being generalizability of this method was investigated by using a
0.06 (5.6%), 0.13 (12.6%), and 0.01 (1.0%), respectively donor-specific trained model to evaluate the deep learning
(Fig. 7). In summary, these results show that our deep derived deformability profile from unsorted donor RBCs
learning model is a robust and generalizable for classifying (Fig. 7). The RS obtained by this approach showed strong
the deformability of RBCs acquired and processed separately agreement with microfluidic sorting, deviating by 1.0–12.6%
from the training dataset. from the microfluidic measurement, which is comparable to
previously reported variability of microfluidic ratchet
Discussion measurement (13.8% from five independent measurements).9
We used saliency maps to confirm that cellular features,
This study developed a deep learning method to assess RBC not imaging artifacts, were used by the convolutional neural
deformability directly from microscopy images in order to network for RBC classification. The saliency maps in Fig. 5
avoid the use of complex and time-consuming physical indicate that cell surface features, especially morphological
measurements. By sorting RBCs into fractions based on contours, are the main characteristics used for deformability
deformability and imaging cells from each fraction, we were prediction. Other deep learning models have been used to

Fig. 7 Generalization of the deep learning network applied to unsorted cell images (light lines) compared to microfluidic-derived deformability
cumulative distributions (dark lines) for donor 2 (A), 3 (B), and 4 (C). The model generalizes well on the unsorted datasets, illustrated by similar
cumulative distribution plots and rigidity scores between the ground truth microfluidic results and the deep learning model tested on an unsorted
RBC sample.

34 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

classify RBCs based on microscopy imaging, including those Methods

related to malaria disease status38 and the RBC storage RBC sample collection and preparation
lesion.37 Both malaria infection58 and storage9,67,68 are
associated with the loss of RBC deformability, suggesting that This study was approved by the University of British
the morphological features detected by these models also Columbia Clinical Research Ethics Board (UBC REB# H19-
correspond to deformability changes. Here, we confirm that 01121) and the Canadian Blood Services Research Ethics
deep learning models can indeed infer cell deformability Board (CBS REB# 2019-029). Blood samples were collected
from morphological features obtained by imaging. from donors following informed consent. Donors who self-
Assessing RBC deformability using microscopy imaging identified as healthy and were between the ages of 18–70
and deep learning has several important advantages over provided fresh RBCs in citrate tubes (n = 3), stored in an
physical measurement. First, although the throughput of Eppendorf tube (n = 1), or stored in blood bags (n = 6)
microfluidic devices used to measure RBC deformability is (Table 3). Donors were diverse in terms of blood type and sex
higher than other methods, the throughput of this process (Table 3).
Blood sample components were separated by centrifuging
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

is limited by the high deformability of RBCs, which requires

the application of small and precisely controlled stress and at 3900 rpm for 8 minutes at room temperature. Plasma
sufficient time to observe a response. In contrast, using supernatant and leukocyte buffy coat were removed and
machine learning to infer RBC deformability allows for disposed. The RBC pellet was resuspended and washed three
evaluation of cells densely seeded within imaging wells at a times using Hanks balanced salt solution (HBSS, Gibco) with
rate of ∼1500 cells imaged per minute. Second, a major 0.2% Pluronic solution (F127, MilliporeSigma) by
barrier to physical measurement of RBC deformability is centrifuging at 1800 rpm for 5 minutes. After all supernatant
that these methods are often difficult, time consuming, and and leukocytes are removed, the RBC pellet was diluted to
require specialized equipment.34,63,64 In contrast, 1% hematocrit in HBSS + 0.2% Pluronic for infusion into the
microscope systems are ubiquitous in both research and microfluidic device.
clinical laboratories. Finally, while there are several methods
available for physical measurement of RBC deformability, Microfluidic ratchet device manufacture
there are no accepted standards by which to compare The manufacture of the microfluidic devices has been
studies. Here, we use the microfluidic ratchet mechanism to described previously.55,58 The master device mold was created
sort cells, which can be calibrated by sorting size-specific using photolithographic microfabrication and was used to
microbeads to provide a measurement standard (Fig. S1†). create a secondary master polyurethane mold fabricated out
Therefore, the universality of microscope systems could of Smooth-Cast urethane resin (Smooth-Cast ONYX SLOW,
offer an approach to standardize RBC deformability Smooth-On) as described here.65 Single-use microfluidic
measurements in order to extend studies across multiple ratchet devices were molded from the secondary master
centers. In summary, RBC assessment by machine learning using PDMS silicone (Sylgard-184, Ellsworth Adhesives)
is more accessible, simpler to perform, and more mixed at a 10 : 1 ratio with the PDMS curing agent (Sylgard-
standardizable compared to physical RBC deformability 184, Ellsworth Adhesives). The PDMS molded devices were
measurements. then cured for two hours at 65 °C. The cured PDMS devices
To our knowledge, this work describes the first instance of were removed from the molds and manually punched with
employing deep learning to predict RBC deformability. Deep 0.5 and 3.0 mm hole punches (Technical Innovations). A thin
learning has been used previously to characterize other PDMS silicone (RTV 615, Momentive Performance Materials
cellular properties of RBCs. For example, Doan et al.37 trained LLC) layer was manufactured to seal the device's
a deep learning model to classify unlabeled images of stored microstructures. This layer was produced by spin coating
RBCs into seven morpho-types with 77% accuracy, which was uncured PDMS on a 100 mm silicon wafer at 1500 rpm for 1
comparable to 83% agreement in manual classification by minute, then was cured for 2 hours at 65 °C. The Sylgard-184
experts. Other studies trained deep learning models to PDMS microstructure mold was bonded to the RTV 615 thin
identify RBCs from patients with malaria,38–43 sickle cell PDMS layer using air plasma (Model PDC-001, Harrick
disease,44–49 and thalassemia,50–52 based on visually Plasma). Finally, the composite sealed microstructure mold
identifiable changes in RBC morphology. Our application of was then bonded to a 75 × 50 mm glass slide (Corning) using
machine learning in RBC deformability measurement air plasma.
deviates from these previous efforts because cellular features
corresponding to deformability are beyond human
perception. This result expands on our previous study using Microfluidic device operation
deep learning to distinguish between cell lines that lack The mechanism of using microscale funnel constrictions to
human distinguishable features,53 which further supports measure cell deformability and the operation of the
our belief that imperceivable cellular parameters, such as microfluidic device has been described and validated
changes in biophysical or metabolic cell state, may be previously.7,9,10,16,31–33,54–57 The microfluidic ratchet sorting
detectable from cell images using deep learning. device is operated via 4 pressurized fluidic inputs. A

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 35
 View Article Online

Communication Lab on a Chip

horizontal crossflow moves the sample towards the outlets, Intra- and inter-user microbead sorting was consistent (Fig.
while a vertical (relative to Fig. 2A–C) oscillating pressure S1†). The statistical analysis is found in the ESI.†
system squeezes cells through the tapered constrictions and
declogs others unable to pass through. The sorting matrix Image acquisition
of micropores has openings ranging from 1.5 μm to 7.5 μm After microfluidic sorting, sorted RBCs were removed from
(Table 1). The resolution of these bins is limited by the the microfluidic device and transferred to a 96-well flat-
microfabrication process, whereby the resolution of the bottom plate (VWR International, LLC). Samples of sorted
mask is limited to 250 nm while our photolithography cells from each outlet were divided evenly and placed into
wavelength is limited to 340 nm. Therefore, the smallest two separate wells to provide cell images captured with
feature resolution possible with this microfabrication variations in well location distribution and automatically
process is ∼250 nm. Before the RBC sample is infused, the determined imaging parameters (e.g. auto-exposure and auto-
device is buffered with HBSS with 0.2% Pluronic-F127 focus) to produce robust datasets. Full image scans of each
solution through the horizontal crossflow inlet at high well in 40× brightfield were acquired using a Nikon Ti-2E
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

pressure (300 mbar) for 15 minutes. Once the device is inverted microscope and NIS Elements software. Illumination
buffered, 10 μL of HBSS with 0.2% Pluronic-F127 solution for the brightfield images was implemented by using the
is pipetted into each outlet (which are open to atmospheric built-in Ti-2E LED. Gain, exposure, and vertical offset were
pressure) to improve ease of removal of each RBC automatically determined by built-in NIS Elements functions
deformability sample after sorting. The RBC sample for for consistency and to avoid user bias. Components of the
each donor is suspended at 1% hematocrit in HBSS with full image scan were 2424 × 2424 pixel BMP images with 24-
0.2% Pluronic-F127 and then infused into the microfluidic bit depth.
device at 40–45 mbar through the sample inlet. The sample
flows through the constriction matrix via the horizontal Segmentation and augmentation
crossflow pressure (55–60 mbar) and oscillatory pressure of
175 mbar upwards and 162 mbar downwards (relative to Each full scan image was segmented using a custom
Fig. 2B and C). The oscillation cycle of these pressures computer vision segmentation algorithm. Individual cells
occurs over 5 seconds: 4 seconds of upward pressure flow were identified using a watershed algorithm and were
for filtration, then 1 second of downward flow for segmented into 60 × 60 pixel PNG images with 8-bit depth.
declogging. The sorting process throughput is approximately Segmented images with multiple or partial cells were
600 cells per minute; the device is run for 60–90 minutes, manually removed. Resultant single cell images from each
resulting in over 30 000 sorted cells. After the cells are donor were split at an 80 : 20 ratio per class to create separate
sorted through the constriction matrix, they proceed to one training and testing datasets. After database splitting, images
of 12 distinct deformability outlets. The distribution of were augmented by a random multiple of 90° rotation (0°,
sorted cells is determined by capturing images as the cells 90°, 180°, or 270°). This augmentation allowed for the
exit the constriction matrix and counting the cells manually building of balanced training (10 000 images per outlet) and
using ImageJ.66 The distribution can also be determined by testing (2000 images per outlet) datasets for each donor. In
video analysis of the cells travelling through the constriction addition, different lighting conditions were observed
matrix exit channels towards the outlets, or by removing the depending on the location of the cell in the well. By
cells in the outlets by pipetting for counting. Sorted RBCs augmenting the cells by rotation this potential data
suspended in 10 μL of HBSS with 0.2% Pluronic-F127 from confounder is mitigated.
each outlet are removed by pipetting and placed in a 96
well plate (VWR International, LLC) for imaging. Convolution neural network model
A CNN, shown in Fig. 1F, was designed in Python 3.7 using
the Keras library in TensorFlow. The network accepts a
Microbead sorting validation 1-channel input of 60 × 60 pixels. The model begins with a
Microbead sorting validation was conducted to ensure device 256-channel convolution layer with a kernel size of 7 and a
manufacturing and sorting consistency between different stride of 1. Next, the model uses a 2 × 2 max-pooling layer
devices and users. To mimic deformable RBCs, 1.53 μm with a stride of 2. The next layer is a 128-channel
polystyrene beads (Cat #17133, Polysciences Inc.) were convolutional layer with a kernel size of 5 × 5 and a stride of
infused into the microfluidic device at 0.1% concentration in 1, followed by another max-pooling layer of size 2 × 2 with a
HBSS with 0.2% Pluronic F127 and 0.2% TWEEN-20 stride of 2. Next are two 64-channel convolutional layers in
(MilliporeSigma) to prevent bead aggregation. The bead series, each with kernel sizes of 3 and strides of 1. These
solution was run through the microfluidic device for 20 layers were followed by a max-pooling layer of size 2 × 2 with
minutes, images were captured as the beads exited the matrix a stride of 2. Then, the layer outputs were flattened into a
sorting region, and the distribution was determined by 1-dimensional array for connection to the fully connected
manually counting beads in ImageJ. Users 1 and 2 conducted layers. Each of the four convolutional layers were followed by
14 total tests using devices from five different master molds. ReLU activation and batch normalization. Then, three 128-

36 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

node fully connected dense layers, consisting of ReLU of health. The views herein do not necessarily reflect the
activation and 20% dropout, were used by the model to learn views of Canadian Blood Services or the federal, provincial,
on the identified features from the earlier convolutional or territorial governments of Canada.
layers. The network outputs two nodes, one per class, with a
SoftMax (normalized exponential) error function for Ethics approval statement
backpropagation.
This study was approved by the University of British
Training environment Columbia's Clinical Research Ethics Board (UBC REB# H19-
01121) and Canadian Blood Services Research Ethics Board
The segmentation and deep learning software were run on a (CBS REB# 2019-029).
desktop PC operating Windows 10 Pro with an AMD Ryzen 7
5800X 8-core processor running at 3.80 GHz. The computer
used 64.0 GB DDR4 RAM running at 3200 MHz. The graphics
Author contributions
card used was a NVIDIA GeForce RTX 3070. Training and H. M. supervised the study. H. M., E. L., and S. P. D.
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

testing were conducted in Python 3.7.11 utilizing the conceived the idea. E. L., E. I., and K. M. performed the
TensorFlow 2.5.0 library. experimental work. E. L. and M. W. performed the
computational work. All authors wrote the manuscript.
Training
For each execution of the network, training occurred for 25 to Conflicts of interest
80 epochs with stochastic gradient descent optimization and
H. M. is listed as inventors on a patent related to this work.
a learning rate between 0.0001 and 0.1. The appropriate
number of epochs and learning rate were determined
iteratively to find the best combination for training
Acknowledgements
convergence and validation accuracy for each donor/dataset We are grateful to Canadian Blood Services' blood donors
combination. Training concluded after there was no who made this research possible.
improvement in the loss after the past five epochs. A batch
size of 32 and a categorical cross entropy loss function from
References
the Tensorflow Keras library (version 2.5.0) were used. The
error function used for backpropagation was the SoftMax 1 J. Ho, W. J. Sibbald and I. H. Chin-Yee, Crit. Care Med.,
function. Additionally, the model was optimized for training 2003, 31, S687–S697.
accuracy and was validated using five-fold cross validation. 2 R. Huisjes, A. Bogdanova, W. W. van Solinge, R. M.
Schiffelers, L. Kaestner and R. van Wijk, Front. Physiol.,
Testing 2018, 9, 656.
3 L. T. Chen and L. Weiss, Blood, 1973, 41, 529–537.
To verify the accuracy on the five validation folds, testing
4 L. Weiss and M. Tavassoli, Semin. Hematol., 1970, 7,
occurred on 2000 images per outlet from the separate testing
372–380.
dataset. This dataset was split from the training set prior to
5 M. H. A. M. Fens, R. van Wijk, G. Andringa, K. L. van
augmentation, ensuring images used for testing were
Rooijen, H. M. Dijstelbloem, J. T. Rasmussen, K. M. K. de
previously unseen by the network. In addition to deep
Vooght, R. M. Schiffelers, C. A. J. M. Gaillard and W. W. van
learning testing conducted on microfluidic sorted cells,
Solinge, Haematologica, 2012, 97, 500–508.
unsorted cell images from donor 2 (6307 images), 3 (9003
6 F. H. Bosch, J. M. Werre, L. Schipper, B. Roerdinkholder-
images), and 4 (2562 images) were also assessed.
Stoelwinder, T. Huls, F. L. A. Willekens, G. Wichers and
M. R. Halie, Eur. J. Haematol., 2009, 52, 35–41.
Data availability statement 7 Q. Guo, S. J. Reiling, P. Rohrbach and H. Ma, Lab Chip,
The data that support the findings of this study are available 2012, 12, 1143–1150.
on request from the corresponding author. 8 J. G. G. Dobbe, M. R. Hardeman, G. J. Streekstra, J. Strackee,
C. Ince and C. A. Grimbergen, Blood Cells, Mol., Dis.,
Funding statement 2002, 28, 373–384.
9 E. Islamzada, K. Matthews, Q. Guo, A. T. Santoso, S. P. Duffy,
This work was supported by grants from the Canadian M. D. Scott and H. Ma, Lab Chip, 2020, 20, 226–235.
Institutes of Health Research (322375, 362500, 414861), 10 K. Matthews, M.-E. Myrand-Lapierre, R. R. Ang, S. P. Duffy,
Natural Sciences and Engineering Research Council of M. D. Scott and H. Ma, J. Biomech., 2015, 48, 4065–4072.
Canada (538818-19, 2015-06541), MITACS (K. M. IT09621), 11 G. J. Streekstra, J. G. G. Dobbe and A. G. Hoekstra, Opt.
and the Canadian Blood Services Graduate Fellowship Express, 2010, 18, 14173.
Program (E. I.), which is funded by the federal government 12 G. J. Streekstra, A. G. Hoekstra, E.-J. Nijhof and R. M.
(Health Canada) and the provincial and territorial ministries Heethaar, Appl. Opt., 1993, 32, 2266.

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 37
 View Article Online

Communication Lab on a Chip

13 A. M. Forsyth, J. Wan, W. D. Ristenpart and H. A. Stone, 38 S. C. Kalkan and O. K. Sahingoz, in 2019 Scientific Meeting on
Microvasc. Res., 2010, 80, 37–43. Electrical-Electronics & Biomedical Engineering and Computer
14 S. S. Lee, Y. Yim, K. H. Ahn and S. J. Lee, Biomed. Science (EBBT), IEEE, Istanbul, Turkey, 2019, pp. 1–4.
Microdevices, 2009, 11, 1021–1027. 39 J. Hung, A. Goodman, S. Lopes, G. Rangel, D. Ravel, F. T. M.
15 Y. Katsumoto, K. Tatsumi, T. Doi and K. Nakabe, Int. J. Heat Costa, M. Duraisingh, M. Marti and A. E. Carpenter, CoRR,
Fluid Flow, 2010, 31, 985–995. 2019.
16 Q. Guo, S. Park and H. Ma, Lab Chip, 2012, 12, 2687. 40 Z. Liang, A. Powell, I. Ersoy, M. Poostchi, K. Silamut, K.
17 M. Lekka, M. Fornal, B. Wizner, T. Grodzicki and J. Styczen, Palaniappan, P. Guo, M. A. Hossain, A. Sameer, R. J. Maude,
Biorheology, 2005, 42, 307–317. J. X. Huang, S. Jaeger and G. Thoma, in 2016 IEEE
18 R. Agrawal, T. Smart, J. Nobre-Cardoso, C. Richards, R. International Conference on Bioinformatics and Biomedicine
Bhatnagar, A. Tufail, D. Shima, P. H. Jones and C. Pavesio, (BIBM), IEEE, Shenzhen, China, 2016, pp. 493–496.
Sci. Rep., 2016, 6, 15873. 41 F. Yang, M. Poostchi, H. Yu, Z. Zhou, K. Silamut, J. Yu, R. J.
19 J. Liu, F. Zhang, L. Zhu, D. Chu and X. Qu, Opt. Commun., Maude, S. Jaeger and S. Antani, IEEE J. Biomed. Health
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

2019, 442, 56–59. Inform., 2020, 24, 1427–1438.

20 J. P. Shelby, J. White, K. Ganesan, P. K. Rathod and D. T. 42 A. Vijayalakshmi and B. Rajesh Kanna, Multimed. Tools Appl.,
Chiu, Proc. Natl. Acad. Sci. U. S. A., 2003, 100, 14618–14622. 2020, 79, 15297–15317.
21 T. Herricks, M. Antia and P. K. Rathod, Cell. Microbiol., 43 Y. Dong, Z. Jiang, H. Shen, W. David Pan, L. A. Williams,
2009, 11, 1340–1353. V. V. B. Reddy, W. H. Benjamin and A. W. Bryan, in 2017
22 S. C. Gifford, J. Derganc, S. S. Shevkoplyas, T. Yoshida and IEEE EMBS International Conference on Biomedical & Health
M. W. Bitensky, Br. J. Haematol., 2006, 135, 395–404. Informatics (BHI), IEEE, Orland, FL, USA, 2017, pp. 101–104.
23 K. Matthews, S. P. Duffy, M.-E. Myrand-Lapierre, R. R. Ang, 44 K. de Haan, H. Ceylan Koydemir, Y. Rivenson, D. Tseng, E.
L. Li, M. D. Scott and H. Ma, Integr. Biol., 2017, 9, 519–528. Van Dyne, L. Bakic, D. Karinca, K. Liang, M. Ilango, E.
24 M.-E. Myrand-Lapierre, X. Deng, R. R. Ang, K. Matthews, Gumustekin and A. Ozcan, NPJ Digit. Med., 2020, 3, 76.
A. T. Santoso and H. Ma, Lab Chip, 2015, 15, 159–167. 45 L. Alzubaidi, M. A. Fadhel, O. Al-Shamma, J. Zhang and Y.
25 J. M. Kwan, Q. Guo, D. L. Kyluik-Price, H. Ma and M. D. Duan, Electronics, 2020, 9, 427.
Scott, Am. J. Hematol., 2013, 88, 682–689. 46 H. A. Abdulkarim, M. A. Abdul Razak, R. Sudirman and N.
26 Q. Guo, S. P. Duffy, K. Matthews, A. T. Santoso, M. D. Scott Ramli, IAES Int. J. Artif. Intell., 2020, 9, 221.
and H. Ma, J. Biomech., 2014, 47, 1767–1776. 47 M. Xu, D. P. Papageorgiou, S. Z. Abidi, M. Dao, H. Zhao and
27 T. Wu, Q. Guo, H. Ma and J. J. Feng, Theor. Appl. Mech. Lett., G. E. Karniadakis, PLoS Comput. Biol., 2017, 13, e1005746.
2015, 5, 227–230. 48 M. Zhang, X. Li, M. Xu and Q. Li, in Medical Image
28 H. Bow, I. V. Pivkin, M. Diez-Silva, S. J. Goldfless, M. Dao, Computing and Computer Assisted Intervention – MICCAI 2018,
J. C. Niles, S. Suresh and J. Han, Lab Chip, 2011, 11, 1065. ed. A. F. Frangi, J. A. Schnabel, C. Davatzikos, C. Alberola-
29 A. Adamo, A. Sharei, L. Adamo, B. Lee, S. Mao and K. F. López and G. Fichtinger, Springer International Publishing,
Jensen, Anal. Chem., 2012, 84, 6438–6443. Cham, 2018, pp. 695–702.
30 A. T. Santoso, X. Deng, J.-H. Lee, K. Matthews, S. P. Duffy, E. 49 M. Zhang, X. Li, M. Xu and Q. Li, IEEE J. Biomed. Health
Islamzada, S. M. McFaul, M.-E. Myrand-Lapierre and H. Ma, Inform., 2020, 24, 3095–3102.
Lab Chip, 2015, 15, 4451–4460. 50 Y.-H. Lin, K. Y.-K. Liao and K.-B. Sung, J. Biomed. Opt.,
31 Q. Guo, S. M. McFaul and H. Ma, Phys. Rev. E: Stat., 2020, 25(11), DOI: 10.1117/1.JBO.25.11.116502.
Nonlinear, Soft Matter Phys., 2011, 83, 051910. 51 D. A. Tyas, S. Hartati, A. Harjoko and T. Ratnaningsih, IEEE
32 S. M. McFaul, B. K. Lin and H. Ma, Lab Chip, 2012, 12, 2369. Access, 2020, 8, 69849–69860.
33 E. S. Park, C. Jin, Q. Guo, R. R. Ang, S. P. Duffy, K. Matthews, 52 S. Purwar, R. Tripathi, R. Ranjan and R. Saxena, in 2021 11th
A. Azad, H. Abdi, T. Todenhöfer, J. Bazov, K. N. Chi, P. C. International Conference on Cloud Computing, Data Science &
Black and H. Ma, Small, 2016, 12, 1909–1919. Engineering (Confluence), IEEE, Noida, India, 2021, pp. 410–
34 S. Shin, J. X. Hou, J. S. Suh and M. Singh, Clin. Hemorheol. 415.
Microcirc., 2007, 37, 319–328. 53 S. Berryman, K. Matthews, J. H. Lee, S. P. Duffy and H. Ma,
35 B. Blasi, A. D'Alessandro, N. Ramundo and L. Zolla, Commun. Biol., 2020, 3, 674.
Transfus. Med., 2012, 22, 90–96. 54 S. M. McFaul, B. K. Lin and H. Ma, Lab Chip, 2012, 12, 2369.
36 M. Bardyn, B. Rappaz, K. Jaferzadeh, D. Crettaz, J.-D. Tissot, 55 Q. Guo, S. P. Duffy, K. Matthews, E. Islamzada and H. Ma,
I. Moon, G. Turcatti, N. Lion and M. Prudent, J. Geophys. Sci. Rep., 2017, 7, 6627.
Res. Space Physics, 2017, 15, 239–248. 56 Q. Guo, S. P. Duffy, K. Matthews, X. Deng, A. T.
37 M. Doan, J. A. Sebastian, J. C. Caicedo, S. Siegert, A. Roch, Santoso, E. Islamzada and H. Ma, Lab Chip, 2016, 16,
T. R. Turner, O. Mykhailova, R. N. Pinto, C. McQuin, A. 645–654.
Goodman, M. J. Parsons, O. Wolkenhauer, H. Hennig, S. 57 Q. Guo, S. P. Duffy and H. Ma, in Microtechnology for Cell
Singh, A. Wilson, J. P. Acker, P. Rees, M. C. Kolios and A. E. Manipulation and Sorting, ed. W. Lee, P. Tseng and D. Di
Carpenter, Proc. Natl. Acad. Sci. U. S. A., 2020, 117, Carlo, Springer International Publishing, Cham, 2017, pp.
21381–21390. 225–254.

38 | Lab Chip, 2022, 22, 26–39 This journal is © The Royal Society of Chemistry 2022
 View Article Online

Lab on a Chip Communication

58 Q. Guo, S. P. Duffy, K. Matthews, X. Deng, A. T. Santoso, E. 63 J. A. Sebastian, M. C. Kolios and J. P. Acker, Transfus. Apher.
Islamzada and H. Ma, Lab Chip, 2016, 16, 645–654. Sci., 2020, 59, 103020.
59 O. Ronneberger, P. Fischer and T. Brox, in Medical Image 64 W. Li, B. Zhu, Y. Cai, Z. Wu, L. Sun and H. Yang, Int. J. Adv.
Computing and Computer-Assisted Intervention – MICCAI Manuf. Technol., 2019, 105, 4919–4928.
2015, ed. N. Navab, J. Hornegger, W. M. Wells and A. F. 65 S. P. Desai, D. M. Freeman and J. Voldman, Lab Chip,
Frangi, Springer International Publishing, Cham, 2015, pp. 2009, 9, 1631.
234–241. 66 C. T. Rueden, J. Schindelin, M. C. Hiner, B. E. DeZonia, A. E.
60 A. Krizhevsky, I. Sutskever and G. E. Hinton, in Advances in Walter, E. T. Arena and K. W. Eliceiri, BMC Bioinf., 2017, 18,
Neural Information Processing Systems, ed. F. Pereira, C. J. C. 529.
Burges, L. Bottou and K. Q. Weinberger, Curran Associates, 67 E. Islamzada, K. Matthews, E. Lamoureux, S. P. Duffy, M. D.
Inc., 2012, vol. 25. Scott and H. Ma, Transfusion, 2021, in press.
61 L. Shamir, J. Microsc., 2011, 243, 284–292. 68 E. Islamzada, K. Matthews, E. S. Lamoureux, S. P. Duffy,
62 K. Simonyan, A. Vedaldi and A. Zisserman, 2014, M. D. Scott and H. Ma, eJHaem, 2021, DOI: 10.1002/jha2.343,
Published on 01 December 2021. Downloaded on 11/15/2022 10:08:47 AM.

arXiv:1312.6034 [cs]. in press.

This journal is © The Royal Society of Chemistry 2022 Lab Chip, 2022, 22, 26–39 | 39