Automated Detection of Sickled Erythrocytes in Peripheral Blood Smear Images Using YOLO-Based Object Detection

Abstract

Abstract Sickle cell disease (SCD) is characterized by abnormal red blood cell (RBC) morphology that drives hemolysis and microvascular complications, yet manual smear evaluation remains subjective and difficult to standardize in heterogeneous fields. In this work, we present a clinically grounded and reproducible framework for cell-level detection of sickled erythrocytes using You Only Look Once (YOLO)-based object detection. The pipeline integrates controlled sample dilution before imaging to reduce cell crowding, an overlapping tiling strategy to preserve morphological detail, and a standardized annotation protocol to minimize ambiguity in borderline cells. Models were trained on peripheral blood smear images from both SCD patients across distinct clinical states, including steady-state and vaso-occlusive crisis. Our results show that YOLO11 achieved the highest performance of mAP50-95 = 0.857, with minimal inter-class confusion. Error analysis indicates that most misclassification arises from background regions and annotation-driven exclusions, highlighting the influence of data quality and labeling strategy on model performance. These findings demonstrate accurate cell-level detection performance of our method in complex smear images and establish a scalable framework for automated RBC morphological analysis, with potential relevance for disease monitoring and deployment in resource-limited settings pending further validation.