Full pixellevel annotation of 3D medical volumes is laborious and costly in clinical practice, particularly in multilesion scenarios where experts must delineate numerous heterogeneous lesions. Consequently, especially for rapid screening, annotators often label only a subset of lesion instances, leaving many unlabeled due to time constraints. In this paper, we formulate such a weak supervision protocol that relies on only a few fully labeled lesion instances and sparse background scribbles, reflecting this clinical reality. Conventional weakly supervised methods, such as scribble, point, or bounding box-based approaches, assume at least partial annotation or location cues for all lesions and thus fail under this protocol. To address this, we propose PIASeg, a meta-learning framework that iteratively refines labels by correcting a subset of highly confident foreground predictions. To further enhance correction reliability, class-specific prototypes are constructed to filter out inconsistent pseudo-labels via feature similarity calculation. Prototype representations are optimized with both contrastive and diversity objectives to ensure robust and rich representations. Extensive experiments on three public 3D lesion datasets, i.e., LiTS, ISLES22, and MS, demonstrate PIASeg's superiority against state-of-the-art baselines. In particular, even when only one lesion per volume is annotated, it still maintains superior segmentation accuracy with such extremely-limited supervision. Code is available at https://github.com/innocence0206/PIASeg.
Guo et al. (2026) studied this question.