In computational pathology, whole-slide images are often labeled at the bag level (e.g., slide-level diagnosis), while instance-level annotations (e.g., tumor regions) are costly and difficult to obtain. Weakly supervised multiple instance learning (MIL) is a common framework for this setting.

The new paper on arXiv introduces ProtoPur, a method that uses prototype-purity pseudo-labeling to improve instance-level learning. The abstract describes the problem and the proposed approach, though details on implementation and evaluation are not included in the available text.

The significance is that reliable instance-level supervision could be derived from weak labels, potentially reducing the need for manual annotation. However, the abstract is truncated, so we cannot comment on experimental results or comparisons.