Objective assessment of stored blood quality by deep learning.

for Innovation, Canadian Blood Services, and 8 (bags CE44 to CE52) at the Transfusion Center of the University Hospital of Geneva, Switzerland. The Canadian Blood Services Research Ethics Board approved (Protocol #nc0058) the collection of the blood products used in this study that were obtained from volunteer, healthy blood donors, who provided written, informed consent. The utilization of blood samples from healthy donors for research was approved by the Ethical Committee of the University Hospital of Geneva. As this was general approval for the use of blood samples for nondiagnostic anonymized research signed by all donors, there is no specific ethical committee approval number. Written informed consent was received from participants, and samples were anonymized prior to inclusion in the study. Further details about sample protocols have been described in Pinto et al. (25) and in SI Appendix, Fig. S1. An additional (third) dataset comprised of hemolytic and IFC measurements of 20 red cell units sampled at 5 storage durations (total 100 data points) were collected at the Blood for Research Facility (netCAD, Vancouver, Canada) and shipped to Canadian Blood Services in Edmonton, Alberta for testing (Fig. 4). The sample preparation protocol for IFC was similar to that of the other Canadian samples. Samples were then analyzed at the University of Alberta Faculty of Medicine and Dentistry Flow Cytometry Facility. One sample in this batch showed an elevated hemolysis levels compared to the others (Fig. 4A). It is unlikely that this was due to bacterial contamination as no common visual indicators of bacterial contamination were present and the hemolysis levels, although higher than the other samples, are still acceptable at day 42 (0.8%). Additionally, a review of the IFC images themselves at two time points did not reveal any significant presence of bacteria. This outlier is therefore more likely caused by donor factors that make this unit more susceptible to hemolysis; this could not be confirmed as the sample was not available for further investigation. IFC Data Acquisition and IDEAS Analysis. For each sample, 5 μL of red cell concentrate were suspended in 200 μL of PBS (magnesium and calcium-free) in a 1.5-mL low-retention microfuge tube (Sigma T4816-250A). Samples were placed on an Amnis ImageStreamX Mark II (Amnis, EMD Millipore), five laser two-camera system (ASSIST calibrated) with a brightfield area lower limit of 50 m2 used to eliminate debris and speed beads. Channels 1, 9 (brightfield), and 12 (dark-field) were used to capture 100,000 brightfield/ darkfield RBC images per sample using the low-speed/high-sensitivity settings at 60× magnification (0.9 numerical aperture, 0.33 per square pixel resolution, 40-μm field-of-view, 2.5-μm depth-of-field). The IFC measurements were repeated for each scheduled time point throughout the blood storage. The instrument-associated analysis software IDEAS v6.2 was used to preliminarily process the acquired IFC data to remove out-of-focus cells, artifacts, debris, and clumped objects, as previously described (25–27). Images of in-focus single cells were then used for manual annotation and downstream deep-learning analysis. Brightfield and darkfield images were exported in .CIF or .TIF formats. Darkfield images were ignored for the final results shown in this study. Ground Truth Annotation. For the supervised machine learning procedure, each RBC was manually annotated by assigned human annotators, in consultation with an RBC morphology expert. Five annotators with different backgrounds (biologists, engineers, and a hematologist) were tasked to manually label allocated RBCs (see next section) as smooth disc, crenated disc, crenated discoid, crenated spheroid, crenated sphere, smooth sphere, sideview, and undecidable class. The undecidable category includes debris or cells that are blurry, blebbed, or folded, and typically represent artifacts of the testing process (SI Appendix, Fig. S3, bottom row; see also description in figure legend). Brightfield and darkfield images of annotated cells were then exported as .TIF. Data Splitting and Validation Strategy. The overall strategy is schematized in SI Appendix, Fig. S4. Training. Image data from replicate samples of bags A, B, D, E, F, H, CE47, CE49, CE50, and CE52 were pooled together. About 17,000 cells of that pooled dataset were annotated by three different annotators. Two annotators were tasked to annotate images from the same blood bags, but different individual cells from them; one annotated cells with an even object index and the other, cells with an odd object index. Finally, one additional annotator reviewed every cell individually and flagged dubious annotation mistakes for correction or removal. Test 1. A class-balanced set of ∼1,500 cells pooled from bags C, G, and I (SI Appendix, Fig. S3) were selected to test interobserver variation and labeling replicability between the five annotators; that is, each individual was tasked to label the exact same cells using an in-house web application (SI Appendix, Fig. S12). Test 3 (morphology). Image sets randomly sampled from (unpooled) bags CE47 and CE49 were used to test the robustness of the trained neural network on imbalanced data. During and after Tests 1 and 2, if suboptimal settings were Fig. 4. Generalizability of SMI to blood samples from a third facility. (A) Additional data for comparison of SMI (as developed in this report) and conventional hemolysis scores of 20 red cell units sampled at five storage durations were analyzed at Canadian Blood Services in Edmonton, Alberta. (B) Hemolytic scores based on the standard physiological hemolysis tests for the collected red cell units. Sample 6 showed an elevated level of hemolysis from day 3 to day 42 (deeper red shades in the table, blue diamonds in B–D), which is likely due to donor factors (Materials and Methods). This data point is therefore marked as blue in the data plots but excluded from statistics. (C) SMI scores by weakly supervised learning of the corresponding red cell units. (D) The correlation between hemolysis and SMI scoring systems. Coefficient of determination R2 = 0.5833. Shaded bands around the regression line display the 95% confidence interval for the regression estimate. With the inclusion of elevated hemolyzed sample (sample 6, shown as blue diamonds), the coefficient of determination R2 is 0.2520, likely because the current neural network was not trained to tolerate certain confounding factors such as donor factors that lead to unusually high hemolysis levels. Doan et al. PNAS | September 1, 2020 | vol. 117 | no. 35 | 21387 BIOPHYSICS AND COMPUTATIONAL BIOLOGY Downloaded at Univ of Connecticut on March 7, 2021 detected, retraining of the supervised and weakly supervised models were allowed and optimization with improved parameters was implemented until the models were satisfactorily considered final. Once finalized, no further changes to the model weights were allowed and only a single inference was done on the hold-out test sets. Tests 3 and 4 (physiology). In particular, bags CE47, CE48, and CE49 have parallel data for both morphological (assayed by an IFC) and physiological (assayed by hemolysis test) assessments. Physiological readouts were used as a means to validate conclusions drawn by morphological findings. Tests 2 and 4. More than 20,000 annotated cells of bags C, G, I, J, CE44, CE45, CE48, and CE51 were kept held-out during the development and optimization of the machine learning algorithms. These data were unlocked only when all machine learning models were final. The prediction on this heldout data were computed a single time, immediately before the submission of the report for the final validation of the trained models. Supervised Deep Learning. Protocols for image preprocessing and deeplearning training of the supervised classification are similar to our previously established label-free imaging flow cytometry machine vision framework (61). In brief, the input images were contrast-stretched channel-wise and resized to 48 × 48 pixels by cropping or padding. To counter illumination variations in image inputs, the data were zero-centered using channelwise mean subtraction and augmentation was implemented, such as random combinations of horizontal or vertical flips, horizontal, or vertical shifts (up to 50% of the image size), and rotations up to 180°. We implemented a ResNet50 architecture (62) (SI Appendix, Fig. S13), with categorical crossentropy as the loss function and accuracy as the performance metric. The model was compiled using the Adam optimizer with a learning rate of 0.0001. The learning rate was reduced by a factor of 10 when the validation loss failed to improve for 10 consecutive epochs. The model was trained for a maximum of 512 epochs, although early stopping generally terminated training before 200 epochs when there is no improvement in the validation loss after 50 consecutive epochs, as detailed in Doan et al. (61). Training and validation data were randomly undersampled per blood unit across cell types to create a balanced dataset. Eighty percent of sampled data were assigned to the training dataset, with the remaining 20% assigned to internal validation of the model during its training. Prediction metrics included recall, precision, F1-score, and weighted acc