WsiSSM: A Weakly Supervised Subset-Matching Framework for Unified Classification and Segmentation of Histopathology Whole Slide Images

Multiple instance learning (MIL) has been increasingly used to analyze histopathology whole slide images (WSIs) and has achieved accurate diagnostic classification and a certain degree of automatic segmentation capability. However, segmentation performance without annotation remains suboptimal due to the immense size of WSIs, despite its essential role in assisting pathologists’ diagnosis. Moreover, existing approaches treat classification and segmentation as independent tasks, with segmentation primarily serving as an interpretability section. To address this limitation, we propose a novel weakly supervised learning algorithm that unifies classification and segmentation within a framework. Specifically, we explore the interconnections between patches and introduce class activation maps in WSI analysis through the proposed Sparse-CAM backbone to activate the segmentation capability, and further enhance feature representation using the proposed subset-matching module to improve both classification and segmentation performance. The proposed method demonstrated superior performance in diagnosis and achieved state-of-the-art performance in segmentation compared to other MIL methods across multiple datasets. In addition, it offered high-precision segmentation that is sensitive to micro-tumor regions without pixel-level annotations. The code will be publicly available on GitHub upon acceptance.