Deep ensemble learning enables highly accurate classification of stored red blood cell morphology

Abstract

Changes in red blood cell (RBC) morphology distribution have emerged as a quantitative biomarker for the degradation of RBC functional properties during hypothermic storage. Previously published automated methods for classifying the morphology of stored RBCs often had insufficient accuracy and relied on proprietary code and datasets, making them difficult to use in many research and clinical applications. Here we describe the development and validation of a highly accurate open-source RBC morphology classification pipeline based on ensemble deep learning (DL). The DL-enabled pipeline utilized adaptive thresholding or semantic segmentation for RBC identification, a deep ensemble of four convolutional neural networks (CNNs) to classify RBC morphology, and Kalman filtering with Hungarian assignment for tracking changes in the morphology of individual RBCs over time. The ensembled CNNs were trained and evaluated on thousands of individual RBCs from two open-access datasets previously collected to quantify the morphological heterogeneity and washing-induced shape recovery of stored RBCs. Confusion matrices and reliability diagrams demonstrated under-confidence of the constituent models and an accuracy of about 98% for the deep ensemble. Such a high accuracy allowed the CNN ensemble to uncover new insights over our previously published studies. Re-analysis of the datasets yielded much more accurate distributions of the effective diameters of stored RBCs at each stage of morphological degradation (discocyte: 7.821 ± 0.429 µm, echinocyte 1: 7.800 ± 0.581 µm, echinocyte 2: 7.304 ± 0.567 µm, echinocyte 3: 6.433 ± 0.490 µm, sphero-echinocyte: 5.963 ± 0.348 µm, spherocyte: 5.904 ± 0.292 µm, stomatocyte: 7.080 ± 0.522 µm). The effective diameter distributions were significantly different across all morphologies, with considerable effect sizes for non-neighboring classes. A combination of morphology classification with cell tracking enabled the discovery of a relatively rare and previously overlooked shape recovery of some sphero-echinocytes to early-stage echinocytes after washing with 1% human serum albumin solution. Finally, the datasets and code have been made freely available online to enable replication, further improvement, and adaptation of our work for other applications.

Big Data in Transfusion Medicine and Artificial Intelligence Analysis for Red Blood Cell Quality Control

“…To determine the RBC morphology quickly and unbiasedly, machine learning approaches are used nowadays, similar to other aspects in hematology and transfusion medicine [24,84,85]. In the context of RBCs, artificial neural networks and deep learning-based techniques have been used to assess cell phenotypes both in stasis [57,82,[86][87][88] and during deformation [71,83,[89][90][91]. Kim et al [82] employed a generative adversarial network to evaluate RBC phenotypes based on phase images obtained by digital holographic microscopy at rest.…”

Advancing precision single-cell analysis of red blood cells through semi-supervised deep learning using database of patients with post-COVID-19 syndrome

“…The availability of labeled data is also a significant hurdle. Presently used data collection methods are often based on microscopy 21,32,37 or custom-made flow devices 15 . Supervised DL models require large amounts of labeled data for training, and obtaining labeled datasets is often labor-intensive and time-consuming, as it involves manual labeling by experts 38 .…”

“…Tasks involving the identification and classification of cells and particles 21,32,36 have been slow and challenging due to the exigent nature of manual gating and the limitations of proprietary code and datasets. Variant cell morphologies and states pose additional challenges for accurate identification, thus necessitating sophisticated computational methods 37,39,40,44 .…”

“…Prior contributions in this field successfully attempted the integration of machine learning algorithms, particularly deep learning, resulting in significant results in classification accuracy and data processing efficiency. However, most of the past works rely on heavy computation models for both classification and localization 26,32,35 Additionally, there has been no mention of poorly focused datasets and attempts to solve the issue of defocus for IFC data. Additionally, past works typically do not optimize different pre-training modes and model configurations 8,22,35,40 .…”

RedTell: an AI tool for interpretable analysis of red blood cell morphology

“…apply Fast R-CNN and Faster R-CNN models to differentiate between 14 and 7 RBC subtypes, respectively. The most recent work suggests a pipeline based on deep ensemble learning for RBC morphology classification (Routt et al, 2023). Such methods result in a higher classification accuracy in multi-class settings.…”