In computational pathology, whole-slide images often contain multiple distinct lesions, yet the available diagnostic label typically reflects only the most severe finding. This creates a weak-supervision problem: a model sees a slide-level label but must learn to identify all relevant regions, not just the dominant one.
The paper formulates this setting as "dominant-label multiple instance learning," where the observed label is the most severe among the instances. The abstract is truncated in the source, so the full method and results are not available, but the framing is clear: the aim is to recover evidence of coexisting lesions even when supervision is limited to the dominant label.
Because this is a single source and the abstract is incomplete, there are no other findings to compare against. The significance lies in naming and formalising a common but under-addressed problem in whole-slide image analysis.