Assaying storage lesion of irradiated red blood cells by deep learning with attention mechanism.

• An improved deep learning with attention mechanism was employed to classify unlabeled RBCs according to morphology, which could be used to assess the storage lesion of stored RBCs with MI and SDCs%. • Compared with the index of MI, SDCs% might be more appropriate for blood quality evaluation under strict requirements for blood transfusion. • The morphological quality of irradiated RBCs has been improved under hypoxic storage. In blood transfusion, irradiation of red blood cells (RBCs) is a widely used treatment method. The storage lesion of RBCs has a significant impact on patient prognosis. How to rapidly evaluate storage lesion and how to extend their storage duration will significantly affect utilization rate. Since the morphology of RBCs is closely related with its physiological function, morphological changes during storage can more intuitively reflect the storage lesion of RBCs. However, manual classification of RBCs was still widely applied in blood transfusion, which may lead to different results in RBCs' classification depending on professional skills. In this paper, we present to automatically classify storage RBCs into nine-subtypes based on an improved YOLOv5s network with attention mechanisms. Under a self-developed microscope with bright-field illumination, RBCs were imaged and treated into labelled dataset according to morphological characteristics. To assess the storage lesion, we presented morphological index (MI) and percentage of smooth disc cells (SDCs) to analyze RBCs' classifications from deep learning. The results indicate that the storage quality characterized by MI is more consistent with clinical standards. Since SDCs% represents the proportion of cells in the blood that can perform oxygen transport, SDCs% is more suitable as an evaluation criterion when there are strict requirements for blood transfusion. Compared with conventional storage condition, hypoxic storage reduces RBCs storage lesion. According to the morphology of RBCs, the quality of irradiated RBCs has been improved under hypoxic storage, which will extend its storage duration. Therefore, hypoxic storage is a potential method for RBC preservation. This is the entire process of detecting red blood cells (RBCs) using deep learning, including RBC classification standards, RBC image acquisition, and data annotation. The left side shows how the datasets were prepared. We recognized and classified RBCs by the improved YOLOV5 network with attention mechanisms. The right side shows the detection results of the 1920×1920 resolution RBC image. Different colored boxes show different morphology of RBCs. [Display omitted] [ABSTRACT FROM AUTHOR]