Dispersion-to-consolidation: Consolidating dispersed semantics via context-aware clustering for whole slide image analysis.

Abstract

Multiple Instance Learning (MIL) has become the dominant paradigm for analyzing histopathology whole slide images (WSIs). However, the inherent spatial heterogeneity of WSIs poses a significant challenge to existing MIL methods. This heterogeneity, where morphologically similar tissue patches often exhibit multifocal distributions across the WSI, hinders the ability of current models to effectively capture long-range spatial dependencies and complex inter-tissue semantic associations. To mitigate these limitations, we propose DisCo, a novel Dispersion-to-Consolidation MIL framework that consolidates dispersed semantics through context-aware clustering for WSI analysis. DisCo operates by iteratively routing and aggregating instances into a compact set of semantic anchors that represent distinct morphological patterns. Specifically, a Cluster Router module dynamically aggregates spatially dispersed yet semantically similar instances into tissue-specific semantic groups, strengthening intra-group feature interactions through dynamic routing and local aggregation. Subsequently, a Cluster Merger module consolidates redundant semantic anchors and captures inter-group associations by learning a dynamic assignment that projects fine-grained anchors into coarser anchors. Extensive experiments on 15 large-scale public cancer datasets across three challenging tasks demonstrate that DisCo consistently outperforms state-of-the-art methods. The source code is available at https://github.com/junjianli106/DisCo.