Fig. 3: Qualitative comparison of segmentation results on two representative gastric adenocarcinoma cases.
From: Geometric multi-instance learning for weakly supervised gastric cancer segmentation
This figure compares our Geo-MIL framework ("Ours") with three key baselines representing different approaches: Patch-WI (classifier activation map), AB-MIL (standard MIL), and HistoGraph (a graph-based method). Left Case: A multi-focal tumor region. Both Patch-WI and AB-MIL fail to identify the full extent of the tumor, producing incomplete and sparse heatmaps. HistoGraph generates a more complete mask but incorrectly merges the two distinct tumor nests. In contrast, our method accurately delineates both regions as separate, complete entities. Right Case: A complex, cribriform glandular structure. Again, the baselines struggle, either capturing only a fraction of the lesion or failing to conform to the intricate boundaries. Our Geo-MIL produces a segmentation mask that is both spatially coherent and anatomically precise, closely matching the underlying pathology.
