A-148 Implementation of YOLO11-RBC Deep Learning Framework for Advanced Erythrocyte Morphological Analysis

Abstract Background Erythrocyte morphological analysis is a critical diagnostic technique in clinical settings. Advanced digital morphology analyzers have enhanced its importance by enabling rapid imaging and quantitative classification of RBCs based on distinct morphological characteristics. However, the algorithms used for identifying abnormal RBC morphology still have room for improvement. Deep learning-based object detection methods offer promising potential in more accurate and simplified RBC morphological analysis. In this study, we introduce a new object detection model framework, YOLO11-RBC, designed for automated classification of abnormal RBCs to support the diagnosis of RBC related diseases. Methods We collected 2106 blood samples with RBC abnormalities from two Chinese institutions. Blood smears were prepared with Mindray SC-120 slide maker and then analyzed using the Mindray MC-80 digital cell morphology analyzer. A total of 9322 digital images of blood smears were captured, with each image containing approximately 100 labeled RBCs. To address class imbalance, we cropped rare RBC types, such as bite cells, blister cells, and Plasmodium-infected RBCs, from 6448 large blood smear images, and performed data augmentation (rotation, deformation, etc.) on the cropped single RBC images. The augmented RBC images were randomly shuffled and stitched into large images. In total, we obtained 520 stitched RBC images, each containing approximately 100 RBCs. Morphologists annotated the stitched RBC images and 1586 original blood smear images based on three main attributes: color (hypochromic cells), shape (schistocytes, helmet cells, bite cells, blister cells, irregularly contracted cells, elliptocytes, ovalocytes, sickle cells, echinocytes, stomatocytes, target cells, teardrop cells, and acanthocytes), and inclusions (e.g., basophilic stippling, Pappenheimer bodies, Howell-Jolly bodies, and Plasmodium). Agglutination and rouleaux formation were also labelled at the image level. The annotated dataset was randomly divided into training and test sets (7:3 ratio). We developed the YOLO11-RBC model, a variant of the YOLO series, optimized for erythrocyte morphological recognition with an improved detection head and loss function. The model*s performance was evaluated on the test set at the individual RBC level and on an external validation set comprising 1288 samples from the third hospital, with samples classified as negative or positive for different RBC abnormalities according to the International Council for Standardization in Haematology (ICSH) guidelines. Results In the test set, the YOLO11-RBC model demonstrated high precision (89.4%–99.0%) and high recall rates (92.6%–99.7%) in the detection of 20 RBC abnormalities, which F1 scores ranging from 92.0% to 99.3%. The harmonic means of these three parameters were 96.4%, 96.7%, and 96.5%, respectively. In the external validation set, the model achieved a sensitivity of 95.0%, a specificity of 98.9%, and an accuracy of 98.7%. Conclusion The YOLO11-RBC model demonstrated strong performance in the classification of abnormal RBCs, with high precision, recall, sensitivity, specificity, and accuracy. These results indicate that the model can accurately and efficiently analyze erythrocyte morphology in clinical samples, making it a valuable tool for automated abnormal RBC classification in clinical practice.