A Benchmark of Eye Anterior Image Semantic Segmentation in Scarcely Labeled Datasets

Semantic segmentation of anterior-segment ophthalmic images is hampered by scarce pixel-wise annotations. We benchmark three data-efficient training paradigms built on a foundation model: (i) parameter-efficient fine-tuning of Segment Anything (SAM) via LoRA, (ii) weakly supervised learning that converts Grad-CAM activation maps from a ResNet-18 classifier into SAM-prompted pseudo-masks, and (iii) few-shot segmentation using feature-density prototypes. Experiments on the public pupillary segmentation Dataset (299 images) evaluate Intersection over Union (IoU), averaging results over 10 runs per configuration. LoRA achieves the best and most stable performance (mean IoU ≈ 0.97 across N=42-210 labeled masks). The weakly supervised pipeline attains competitive accuracy without pixel-wise labels (peak IoU 0.85 N=168) but is sensitive to thresholding. The few-shot method reaches IoU 0.75 with k=20 masks and offers a viable path when only a handful of annotations are available, albeit with higher variance. These findings delineate a practical 'supervision ladder': use LoRA when tens of masks can be curated; prefer weak supervision when only image-level labels exist; and resort to few-shot learning for minimal mask budgets. Overall, coupling foundation models with annotation-efficient training enables clinically useful ophthalmic segmentation under realistic data scarcity.