Abstract
Cataracts remain the leading cause of visual impairment worldwide. We focus on improving the you only look once (YOLO) architecture through targeted optimization to enhance feature extraction. We trained the optimized YOLOv8 detector using 11,274 annotated fundus and anterior segment images. During training, five-fold cross-validation, color magnification, and stochastic weight averaging (SWA) were applied to ensure convergence. In the external test set, the model achieved an F1-score of 98.9% and an mAP50 of 0.995. On an NVIDIA RTX A2000 GPU, the inference speed reached 520 frames per second. Our network enables real-time cataract diagnosis on low-cost GPUs, surpassing previous ResNet- and MobileNet-based benchmarks by ?4% in F1-score and reducing output latency by 68%.