Work

Pathologist-inspired Hierarchical Mamba network for robust BCC subtype prediction across multi-center whole slide images

Year: 2026

Type: article

Abstract: Basal cell carcinoma (BCC) is the most prevalent cutaneous malignancy, with rising incidence worldwide. Accurate distinction between low-risk and high-risk subtypes is critical for guiding surgical strategies, postoperative management, and early therapy. Current multiple-instance learning (MIL) approaches for whole slide images (WSIs) often overlook the spatial continuity of tumor regions, limiting subtype classification performance. We propose the Pathologist-Guided Hierarchical Mamba (PHM), a novel MIL framework designed to mirror the clinical workflow of pathologists. PHM encodes local, regional, and global representations of WSIs, while Adaptive Enhanced Sampling prioritizes high-importance image sequences to improve discriminability and emphasize rare subtypes. The model was trained and validated on a large multi-center BCC dataset collected from three leading Chinese institutions, representing the most extensive subtype classification cohort to date. The proposed PHM framework achieved state-of-the-art performance with 95.25% AUC and 87.07% ACC on internal validation, surpassing S4-Model by +2.90% AUC. External validation confirmed strong generalization (93.36% AUC). Ablation studies demonstrated synergistic effects of hierarchical Mamba, cross-regional attention, and enhanced sampling components. PHM effectively integrates hierarchical slide representation with enhanced global sampling, yielding improved recognition of BCC subtypes. This work demonstrates robust applicability for clinical practice and establishes the most comprehensive framework for subtype-specific BCC diagnosis within the Chinese population. • PHM with cross-regional attention mirrors pathologists’ workflow for BCC subtyping. • Adaptive sampling boosts rare subtype recognition by emulating pathologists’ focus. • Multi-center validation on eight Chinese subtypes ensures clinical applicability.

Source: Biomedical Signal Processing and Control

DOI: 10.1016/j.bspc.2026.110236