Abstract
Weakly supervised segmentation of cancerous regions in whole-slide images (WSIs) is a crucial task in computational pathology, but it is severely hampered by the need for expensive pixel-level annotations. Existing Multiple Instance Learning (MIL) frameworks, while popular, typically fail to produce accurate segmentation masks because they treat WSIs as an unordered 'bag-of-patches', ignoring the critical tissue topology and architectural patterns that define malignancy. In this paper, we address this fundamental limitation by proposing Geometric Multi-Instance Learning (Geo-MIL), a novel graph-based framework that explicitly models the spatial relationships between tissue patches. At the core of our method is a new topological attention mechanism that operates on the WSI graph, learning to identify and prioritize entire diagnostically relevant tissue structures over isolated patch features. Through extensive experiments on three public gastric cancer datasets, we demonstrate that Geo-MIL significantly outperforms a wide array of state-of-the-art baselines, achieving a new benchmark in both segmentation accuracy and classification performance. Our work represents a significant step towards bridging the gap between weak slide-level labels and precise, pixel-level predictions, paving the way for scalable and accurate quantitative analysis in digital pathology.