Objective assessment of stored blood quality by deep learning

Abstract

Stored red blood cells (RBCs) are needed for life-saving blood transfusions, but they undergo continuous degradation. RBC storage lesions are often assessed by microscopic examination or biochemical and biophysical assays, which are complex, time-consuming, and destructive to fragile cells. Here we demonstrate the use of label-free imaging flow cytometry and deep learning to characterize RBC lesions. Using brightfield images, a trained neural network achieved 76.7% agreement with experts in classifying seven clinically relevant RBC morphologies associated with storage lesions, comparable to 82.5% agreement between different experts. Given that human observation and classification may not optimally discern RBC quality, we went further and eliminated subjective human annotation in the training step by training a weakly supervised neural network using only storage duration times. The feature space extracted by this network revealed a chronological progression of morphological changes that better predicted blood quality, as measured by physiological hemolytic assay readouts, than the conventional expert-assessed morphology classification system. With further training and clinical testing across multiple sites, protocols, and instruments, deep learning and label-free imaging flow cytometry might be used to routinely and objectively assess RBC storage lesions. This would automate a complex protocol, minimize laboratory sample handling and preparation, and reduce the impact of procedural errors and discrepancies between facilities and blood donors. The chronology-based machine-learning approach may also improve upon humans' assessment of morphological changes in other biomedically important progressions, such as differentiation and metastasis.

Keywords: cell morphology; deep learning; stored blood quality; weakly supervised learning.

Fig. 1.

Two alternate machine-learning pipelines to assess RBC quality by morphology. The input of the pipeline is single-cell RBC images from imaging flow cytometry. (A) Supervised learning automates the classification of cells into expert-defined categories (SI Appendix, Figs. S1 and S3). The neural network ResNet50 was trained to classify each individual cell into one of the seven morphology classes (smooth disc, crenated disc, crenated discoid, crenated spheroid, crenated sphere, smooth sphere, and side view), as guided by expert annotations of those classes. (B) Weakly supervised learning, by contrast, learns a new quality metric independent of human input, the SMI. The network ResNet50 was trained to identify the storage date of the blood unit a given RBC could belong to, as an auxiliary task. The morphological features extracted by a layer of the network during the training phase can be then used to assign each cell to a point along a continuum from healthy to degraded.

Fig. 2.

Supervised deep learning (Automated Morphology Index) approaches human-level performance for assessing RBC morphology. (A and B) Validation of the supervised deep-learning classifier across two facilities, which include distinct instruments, operators, sample preparation procedures, and donors. Samples were collected independently, with no effort to standardize across the two sites. Most “errors” are in chronologically adjacent categories. Confusion matrices show the prediction of seven categories of RBC morphologies performed by a ResNet50 model (A) trained on the Canadian dataset (n = ∼15,500 cells) and tested on Swiss data (n = ∼25,400 cells) and (B) vice versa; comparable accuracy is seen in both cases. (C) Discrepancies between five human annotators when assigning the exact same cells (n = 1,500) into RBC morphology classes. Detailed analysis of human discrepancies is shown in SI Appendix, Figs. S6 and S10; the average is shown here. (D–F) Validation of the trained supervised models on held-out datasets (Test 2 in SI Appendix, Fig. S4). The held-out datasets were not used in training and were only tested once, immediately before the submission of the work. As is the case for the supervised deep learning model in A and B, most of the errors are in adjacent classes, pointing to inconsistency in human-defined categories (SI Appendix, Fig. S11). Because the accuracy shown in F, 76.7%, is comparable to that between experts (in C, 82.5%), we conclude the trained deep-learning model is roughly as effective as an expert.

Fig. 3.

Data-driven ordering of RBC morphologies by weakly supervised learning allows robust blood quality assessment. (A–C) We discovered a relatively linear progression for major morphological classes of RBCs using features extracted from an intermediate layer (Res4a_ReLU) of the trained weakly supervised model projected into low-dimensional space using a UMAP algorithm. This continuum is observed in 3D (A), 2D (B), or 1D (C) embedding space and interactive 3D-PCA, t-SNE, and UMAP projections of 7,000 representative cells can be explored in a public browser-based tool (ref. 36), select colors and labels for better visualization). Color coding in A, B, and C is consistent, showing that the extracted weakly supervised features place cells along their correct biological progression, from discocytes (smooth discs, crenated discs) to echinocytes (crenated discoids, crenated spheroids, crenated spheres) to spherocytes (smooth sphere). The boxes in A overlap due to continuous transitions between morphology categories, which could not be further resolved by the chosen ResNet50 architecture. The red dotted line in C indicates the threshold in the 1D UMAP, above which RBCs were categorized as unhealthy; this includes most spheroechinocytes (crenated spheroid, crenated spheres, and smooth spheres). The increasing fraction of unhealthy cells (x) over the total number of cells is termed SMI. (D) Distribution of unannotated cells according to the 1D UMAP of weakly supervised features. For each blood unit, deep-learning features were extracted from label-free images of 20,000 cells by a trained weakly supervised neural network. The extracted features were then projected in 1D UMAP space (x axis of each histogram). The shift of distributions from the left to the right as time progresses is clearly visible (more healthy biconcave RBCs are toward the negative end of the x axis; spheroechinocytes are toward the positive end). (E–H) Results from different approaches for evaluating the quality of blood units, with the y axes unified to the same scale. (E) Blood quality according to the proposed SMI. (F) Blood quality according to our automated MI morphology analysis using a fully supervised classifier. (G) Blood quality as assessed by a physiological assay for hemolysis. (H) Blood quality as assessed by expert manual MI morphology analysis (∼4,000 cells per blood unit per time point). (I–L) Pairwise comparisons between proposed machine learning approaches and classic methods for evaluating the quality of red cell units. (I) There is a stronger correlation between the proposed weakly supervised-based quality assessment and hemolytic readouts (coefficient of determination, R² = 0.93) than that of (J) human manual annotations of morphology (R² = 0.67). (K and L) In contrast, the proposed supervised learning-based method shows the opposite trend. The x and y axes of plots I–L were unified to the same scale.

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.