Effective clustering with K-means or efficient cropping with YOLO as a preprocessing step for automated fundus image standardization

This study investigates the effectiveness of K-means clustering and YOLO-based cropping for automated fundus image standardization to enhance optical cup and disc segmentation. The first experiment evaluates K-means clustering’s impact on segmentation performance using the EARDS model, showing that clustered images achieved higher average dice scores compared to non-clustered images. The second experiment employs YOLOv11 for optical disc detection, achieving high performance metrics with a mean average precision close to one. The segmentation model applied to cropped images yielded improved average dice scores for the optical cup and disc, with comparable inference times. While clustering does not significantly enhance segmentation performance, effective cropping is crucial for optimizing model accuracy in fundus image analysis.