Few-Shot Weakly Supervised Segmentation for Retinal Fundus Images using Meta-Learning

Automated segmentation of the optic disc (OD) and the optic cup (OC) in retinal fundus images plays a pivotal role in early glaucoma diagnosis. Many studies have employed deep learning models, but they typically require large-scale densely labeled datasets, which are challenging to obtain. Few-shot weakly supervised segmentation (FWS) emerges as a promising solution, since it requires only a small number of sparsely labeled images. While some studies have explored meta-learning methods for FWS, few to none of them have been applied to retinal fundus images. Thus, we adapt and refine two cutting-edge methods, WeaSeL and ProtoSeg, for multiclass segmentation on multichannel fundus images. We introduce an innovative meta-training strategy to improve data uniformity. Additionally, we develop sparsification techniques for multiclass labels, including points, grids, contours, skeletons, and regions. Multiple scenarios, variations in sparse labels, and the number of shots are evaluated. Our findings suggest that integrating all types of sparse labels in a single meta-training enables the model to learn them more effectively. WeaSeL achieves an Intersection over Union (IoU) of 0.94 for OD and 0.90 for OC using five shots with grid sparse labels. ProtoSeg achieves an IoU of 0.95 for OD and 0.92 for OC using only one shot with contour sparse labels. These results surpass unsupervised domain adaptation approaches and are comparable to fully supervised methods.