University of British Columbia, Mechanical Engineering, Centre for Blood Research
Predicting Red Blood Cell Deformability from Microscopy Images Using Deep Learning
Erik Lamoureux, Matthew Weins, Emel Islamzada, Kerryn Matthews, Hongshen Ma
Introduction Individual Donor Deformability Prediction
Red blood cell (RBC) deformability characterizes the cell’s ability to squeeze through the
microvasculature. Transfusion RBC deformability varies based on donor characteristics
[1,2] and the storage time [3]. More deformable blood increases skin blood flow [4],
haemoglobin levels [5], and is hypothesized to endure for longer periods in the body after
transfusion compared to less deformable cells. The identification and use of long-lasting
transfusion blood would reduce the frequency of needed transfusions for chronic
transfusion patients, reducing their risk of iron overload or other transfusion-related
morbidities. We have developed a convolutional neural network (CNN) approach to
determine blood cell deformability, compared against microfluidic cell sorting, that is
accessible to anyone with an adequate optical microscope.
Hypothesis
We hypothesize that a convolutional neural network model can classify different RBC
deformability levels using optical microscope images compared to microfluidic sorting.
Objectives
• Determine if deformability can be determined using CNNs compared to microfluidic
sorting among: Figure 4: RBC deformability profile prediction for 10 donors compared to microfluidic sorting. R50 rigidity
o a single donor and scores are similar for both methods. Deep learning testing accuracies ranged from 68% (Donor F) to 98%
o multiple donors. (Donor D) with an aggregate average (±SD) of 84±11%.
Microfluidic Ratchet Device Design Combined Donor Deformability Prediction
A B
A
Training on 9 Donors
Testing on 1 Held-Out Donor
Fails Generalization
B
Figure 5: Combined Donor Deformability Prediction. A: Training on 10 donors and testing on held-out subset
of these 10 donors yields comparable results to single donor deformability predictions. B: Training on 9
donors and testing on 1 held-out donor. Method fails with 44% accuracy. The model is not able to generalize
deformability to an unseen donor dataset.
Differentiating Between Donors
Figure 1: Microfluidic Ratchet Device Design. A: An RBC sample is run through a microfluidic A B
deformability sorting device. B: The cells are squeezed through progressively narrower tapered
constrictions. Once the cells reach their constriction limit, they are sorted to a distinct deformability outlet.
RBC deformability determined by the device is the “ground truth” for the machine learning model.
Microfluidic Sorting Results
A B
Figure 6: Differentiating between all 10 donors (all images from deformability outlet 3). A: Model converges
readily in 25 epochs, achieving 93% training accuracy and 90% validation and testing accuracy. B:
Normalized confusion matrix displaying true positive donor predictions along the diagonal. The model more
readily identifies differences between donors than differences in deformability, providing a reason for failure in
deformability generalization (Figure 5B).
Conclusions and Future Directions
Figure 2: Sample microfluidic sorting results from donors with deformable and more rigid blood profiles.
A: Normal distribution of sorted cells; outlets 2 to 5 are used for image classification. B: Cumulative In summary, we developed a deep learning model that accurately determines a RBC’s
distribution plot of sorted cells; the 50% crossover frequency is defined as the donor’s rigidity score. deformability compared with microfluidic deformability sorting. The model works well on
individual donors and with many donors combined but fails to generalize deformability
prediction when provided data from a new donor. This generalization failure is likely caused
Deep Learning Data Pipeline by there being greater inter-donor RBC image variation than inter-donor RBC deformability
commonality. Future work aims to predict RBC storage time from microscopy images due to
storage-related morphological changes.
References
[1] R. Huisjes, A. Bogdanova, W. W. van Solinge, R. M. Schiffelers, L. Kaestner, and R. van Wijk, ‘Squeezing for Life – Properties of
Red Blood Cell Deformability’, Front. Physiol., vol. 9, p. 656, Jun. 2018, doi: 10.3389/fphys.2018.00656.
[2] E. Islamzada, ‘Deformability based sorting of stored red blood cells reveals donor-dependent aging curves’, Lab. Chip, p. 11,
2020.
[3] A. D’Alessandro, G. M. Liumbruno, G. Grazzini, and L. Zolla, ‘Red blood cell storage: the story so far’, Blood Transfus., 2010,
doi: 10.2450/2009.0122-09.
[4] G. Barshtein et al., ‘Deformability of transfused red blood cells is a potent determinant of transfusion-induced change in
Figure 3: Deep Learning Data Pipeline. A: 40x brightfield full well image scans conducted on Nikon recipient’s blood flow’, Microcirculation, vol. 23, no. 7, pp. 479–486, Oct. 2016, doi: 10.1111/micc.12296.
optical microscope. B: Cells are segmented into 60x60 pixel images and augmented by random multiples [5] G. Barshtein, N. Goldschmidt, A. R. Pries, O. Zelig, D. Arbell, and S. Yedgar, ‘Deformability of transfused red blood cells is a
of 90-degree rotations to create balanced classes. C: Schematic of convolutional neural network, similar potent effector of transfusion-induced hemoglobin increment: A study with β-thalassemia major patients: BARSHTEIN et al.’, Am. J.
to AlexNet [6]. Hematol., vol. 92, no. 9, pp. E559–E560, Sep. 2017, doi: 10.1002/ajh.24821.
[6] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘ImageNet classification with deep convolutional neural networks’, Commun. ACM,
vol. 60, no. 6, pp. 84–90, May 2017, doi: 10.1145/3065386.
Contact
Dr. Hongshen Ma, [email protected]
Erik Lamoureux, [email protected]
