Fig. 4: High-fidelity segmentation results of our Geo-MIL framework on two distinct cases.
From: Geometric multi-instance learning for weakly supervised gastric cancer segmentation
This figure showcases the fine-grained accuracy of our model’s predictions. The columns represent the original input patch, the pathologist-annotated ground truth mask, and our model’s final segmentation result. Case A (top row) features a complex, high-grade tumor nest with highly irregular and intricate boundaries. Case B (bottom row) presents a different morphological variant with a more lobulated structure. In both challenging examples, the segmentation mask generated by our model (right column) shows a remarkable concordance with the ground truth (middle column). This demonstrates the model’s ability to learn and precisely delineate complex tumor borders, a key capability for enabling accurate downstream quantitative analyses such as tumor area measurement or invasion front assessment.
