MimicPath: Fine-Grained Mitosis Detection and Prognostic Risk Stratification in Rare Tumor Whole-Slide Images
Abstract:
Mitosis detection in whole-slide images provides key histopathological evidence for prognostic risk stratification in gastrointestinal stromal tumor (GIST), a rare, recurrence-prone mesenchymal tumor. Existing pathology foundation models have shown promising performance in common cancer subtyping and tissue classification, but remain suboptimal for reliable fine-grained mitosis detection in rare tumor whole-slide images (WSIs) with extreme target sparsity and pronounced morphological heterogeneity. To address this challenge, we present MimicPath, a novel vision-language framework that mimics the clinical diagnostic workflow by progressively focusing on mitosis-relevant regions in WSIs. It achieves fine-grained mitosis detection and accurate prognostic risk stratification across multi-center GIST WSIs. Specifically, MimicPath differs from existing approaches in three important ways. (1) We introduce a visual-concept prompted focus (VCPF) module that converts learned disease-specific concepts into spatially aware attention priors, effectively localizing sparse mitosis-relevant regions in gigapixel WSIs. (2) We propose a cross-center decoupling (CCD) module that disentangles domain-agnostic features from domain-specific variations via a dual-branch adaptive gating strategy, enhancing generalization across multiple centers. (3) We design a dual-query parsing (DQP) mechanism that integrates structured clinical reports as textual queries to guide multi-stage visual modulation. It enables clinically grounded inspection and interpretation of suspicious regions, improving the discrimination of subtle mitotic patterns for fine-grained analysis. Extensive evaluations on four in-house GIST cohorts and two public mitosis benchmarks show that MimicPath achieves state-of-the-art performance, underscoring the promise of orchestrating collaboration among generalist foundation models, disease-specific models, and clinicians for intelligent pathology.
