Hierarchical Cross-Scale Attention-Based Multi-Instance Learning for Whole Slide Image Classification

Automated analysis of whole slide images (WSI) integrates image processing algorithms with domain-specific pathophysiological knowledge to assist pathologists in disease diagnosis. Within this domain, deep learning-based WSI classification has emerged as a prominent research focus in recent years. However, existing models often extract features exclusively from a single scale, neglecting rich pathological information present at other scales, resulting in inadequate generalization for certain classifiers. To address this limitation, this paper proposes a Hierarchical Cross-Scale Attention-Based Multi-Instance Learning model (HCSA-MIL) to enhance WSI classification performance. The framework comprises three core modules: (1) Multi-Scale Feature Extraction: Concurrent extraction of patch-level features at ×5, ×10, and ×20 magnification levels; (2) Hierarchical Cross-Scale Attention: Fusion of macroscopic localization information (low-magnification) and microscopic morphological details (high-magnification), explicitly modeling inter-scale complementarity through attention mechanisms; (3) Attentional Feature Aggregation: Adaptive selection and aggregation of discriminative feature embeddings via global attention pooling (GAP) for final classification. Evaluation on the CAMELYON-16 benchmark demonstrates state-of-the-art accuracy of 96.9%, surpassing existing MIL methods by 1.6~7%. This framework exhibits strong potential for automated analysis across diverse WSI applications.