Meta-learners for few-shot weakly-supervised optic disc and cup segmentation on fundus images.

Automatic optic disc (OD) and optic cup (OC) segmentation is essential for glaucoma diagnosis, but limited labeled data remain a major challenge. We address this by developing meta-learners for few-shot weakly-supervised segmentation (FWS). The weak supervision is in the form of sparse labels where only a few pixels are labeled. We significantly improve existing meta-learners by introducing Omni meta-training that enhances data utilization and diversifies the number of shots. We also develop efficient versions that reduce computational costs while maintaining strong performance. In addition, we develop sparsification techniques that simulate customizable and representative scribbles, points, regions, and other sparse labels. Comprehensive evaluations are performed on DRISHTI-GS, REFUGE, and RIM-ONE r3 datasets. We find that Omni and efficient versions outperform the original versions, with the best meta-learner being Efficient Omni ProtoSeg (EO-ProtoSeg). It achieves intersection over union (IoU) scores of 88.15% for OD and 71.17% for OC on the REFUGE dataset using just one sparsely labeled image, outperforming few-shot and semi-supervised methods that require more labeled images. EO-ProtoSeg is also comparable to unsupervised domain adaptation methods, yet much lighter with less than two million parameters and requires no retraining. The results highlight the potential of FWS as a lightweight and effective approach in low-label scenarios.