Weakly Supervised Anterior Lens Surface Segmentation for Cataract Detection

Accurate segmentation of cataract opacities on the anterior lens surface in anterior-segment images could improve early diagnosis and treatment planning in ophthalmology. Automated algorithms could play an important role in addressing the worldwide shortage of clinicians. However, manual annotation of medical images is time-consuming, costly, and often requires expert knowledge. In this study, we propose a weakly supervised learning framework for anterior lens surface segmentation for cataract detection that leverages limited labeled data alongside a larger set of unlabeled images. Our approach integrates class activation maps generated by a deep neural network (trained only with image-level labels) with the output of a foundation segmentation model, minimizing annotation effort. We evaluate the model on a curated dataset of anterior-segment images, demonstrating competitive performance compared to fully supervised baselines. The results suggest that weak supervision can be a viable strategy for scalable and efficient cataract detection, potentially improving access to automated screening tools in resource-limited settings. Notably, the accuracy gains of SAM in Automatic Mask Generator (AMG) mode come with higher inference cost, highlighting a clear accuracy-efficiency trade-off in deployment.