HRGC-YOLO for Urine Sediment Particle Detection in High-Resolution Microscopic Images

Abstract

The automatic detection of urine sediment particle (USP) in microscopy images plays a vital role in evaluating renal and urinary tract diseases. Convolutional neural networks (CNN)-based object detectors have demonstrated remarkable precision in end-to-end detection. However, directly applying CNN-based detectors to high- resolution USP microscopic images poses two major challenges: classification confusion and underutilization of fine-grained information. To address these problems, we present a novel High-Resolution Global Context (HRGC)-YOLO model, which based on YOLOv5m structure and incorporates a global context (GC) block to capture long-range dependencies. Meanwhile, we employ a tile-based detection approach to leverage the uncompressed fine-grained information in high-resolution images. We evaluated the performance of HRGC-YOLO on high-resolution USP datasets from clinic. Compared to YOLOv5m, our HRGC-YOLO network achieved a 4.5% improvement in mAP and outperformed all tested YOLO series models. Our results demonstrate the effectiveness of the proposed method in accurately detecting USPs in high- resolution images.