Detection of crystals and cells in urine sediment content with YOLOv7 segmentation model

In this study, the YOLOv7 segmentation model was utilized to address diagnostic challenges in urine sediment analysis, enabling pixel-level detection of various cellular structures. The model was trained and evaluated on 680 images comprising five classes: red blood cells (RBC), white blood cells (WBC), epithelial cells (Eph), crystals (Crys), and cell populations (Cpop). Unlike traditional bounding box-based methods, YOLOv7 segmentation offers precise identification of overlapping and irregularly shaped particles. Transfer learning and diverse data augmentation techniques were applied to enhance model generalization. The results demonstrate that the YOLOv7 segmentation model achieved high accuracy (mAP@0.50 > 0.97) in clinically significant classes such as RBC, WBC, and Eph, while showing lower performance for structurally complex classes like Cpop and Crys. Mask-based segmentation exhibited notable advantages in dense microscopy fields with overlapping cells. This study indicates that automated segmentation systems can significantly improve diagnostic accuracy in urine sediment analysis and contribute to the advancement of clinical decision support tools.