How a weakly-supervised medical image segmentation model can match or beat full pixel-level supervision by changing the annotation format instead of inventing pseudo labels. The annotator draws an inscribed polygon inside the lesion and an envelope polygon around it, which splits every image into certain foreground, certain background, and an explicitly delimited uncertain annulus, reframing the problem from labelled-versus-unlabelled to certain-versus-uncertain. Two Dice losses trained against the two masks give deliberately contradictory supervision inside the annulus, driving the network toward features invariant to the annotation ambiguity. A classification-guided confidence generator combines thresholded predictive entropy with an auxiliary three-way pixel classifier to decide which annulus pixels are resolvable, and the resolved pixels feed a pixel-wise supervised contrastive loss with a certain-region memory queue and hard-anchor sampling. Covers the staged training schedule, the backbone-agnostic evaluation on ISIC2017 and Kvasir-SEG against scribble-, box-, and fully-supervised baselines, cross-dataset transfer, and the annotation-cost accounting that puts the workload under a fifth of dense labelling.
A computer-vision paper (cs.CV, cs.AI) that attacks the accuracy gap between weakly-supervised and fully-supervised medical image segmentation by changing the annotation format itself rather than by …