Fine-Tuning YOLOv12 for Real-Time Cataract Detection on Edge Devices

Cataracts remain a leading cause of preventable blindness worldwide, emphasizing the need for accurate and scalable early detection. This work proposes a real-time cataract detection system based on the YOLOv12 object detection framework, optimized for deployment in low-resource clinical settings and mobile health applications. A custom dataset of 2,400 annotated ocular images (1,200 cataract, 1,200 normal) was prepared using Roboflow and enhanced through advanced data augmentation techniques to improve generalization.The model incorporates R-ELAN and area-based attention modules for improved spatial feature extraction, along with FlashAttention and mixed-precision training to enhance computational efficiency. Following initial detection, cropped bounding box regions are processed through a secondary classification stage, further improving diagnostic accuracy.Designed for real-time, on-device inference, the system achieves an inference time of 15 ms per image. It reports an F1- score of 83.0%, precision of 85.0%, recall of 81.0%, mAP@0.5 of 86.0%, and mAP@0.5:0.85 of 65.0%. These results confirm the system’s suitability for scalable deployment in telemedicine platforms and remote screening programs, offering a reliable AI-assisted tool for early cataract diagnosis and vision care in underserved populations.