SlideBank Enhances Whole-Slide Image Reasoning for Pathology AI

Whole-slide images (WSIs) present significant challenges for vision-language reasoning due to their gigapixel scale, sparse and heterogeneous diagnostically relevant morphology, and multiple spatial resolutions. Existing WSI models often struggle to semantically access retained information or preserve its connection to original visual evidence after exploration. Researchers have introduced SlideBank, a novel training-free framework designed to address these issues. SlideBank represents each WSI as a persistent, concept-indexed, and spatially grounded evidence bank. It conducts question-independent coarse-to-fine exploration to identify informative regions and multi-scale views, then converts these into explicit morphological observations. Crucially, it grounds pathology signals to their supporting patches and WSI coordinates. During inference, questions are routed to relevant signals and evidence scales, with linked global, regional, and patch evidence integrated through confidence-based cross-level consensus. Experiments on WSI-VQA and SlideBench-BCNB demonstrate that SlideBank significantly improves accuracy with various vision-language models. Furthermore, reusing the same bank across repeated queries achieves over 99% rephrasing consistency and substantially reduces amortized inference costs through persistent evidence reuse.