Robust and Generalizable Detection of Scleral Spots and Dots in Complex Ocular Surfaces With Attention and Mamba‐Enhanced YOLOv5

ABSTRACT Scleral plaque spots and pigmentation dots have significant potential for non‐invasive disease diagnosis. However, the high intra‐class variability in their size, shape, and distribution across individuals poses a critical challenge for automated and robust detection in complex scleral images. This study presents an integrated system that combines sclera imaging and object detection to overcome these challenges. First, a specially designed sclera imaging device is employed to obtain the high quality and shadowless scleral images. Then, an improved YOLOv5 model is proposed, incorporating three key enhancements: a CBAM module to suppress redundant information, a series of Mamba blocks to expand the model's receptive field in scleral images, and a small head strategy to enhance the detection effectiveness on small objects. The proposed detection algorithm was validated on three disparate scleral datasets. The experimental results demonstrate that the proposed method can effectively detect scleral spots and dots, featuring robust and generalizable performance.