Structure-aware generalization for heterogeneous histopathology via prototype-based multiple instance learning.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Accurate cancer diagnosis from whole slide images (WSIs) is hindered by limited annotations, complex tumor structures, and domain shifts.
• Existing multiple instance learning (MIL) methods struggle with generalization and interpretability in computational pathology.

Purpose of the Study:

• To introduce StructMIL, a novel framework for robust and interpretable cancer detection and grading from WSIs.
• To improve the accuracy, generalization, and interpretability of computational pathology models.

Main Methods:

• Developed a structure-aware, prototype-driven MIL framework (StructMIL).
• Integrated graph-based topological priors and histological context.
• Employed prototype-enhanced pooling for stable predictions.
• Implemented a domain-generalization strategy including contrastive alignment, adversarial confusion, and consistency regularization.

Main Results:

• StructMIL achieved state-of-the-art performance on Camelyon16 (breast cancer metastasis detection) and PANDA (prostate cancer Gleason grading).
• Improved cross-center AUC by +3.2% on Camelyon16 (AUC 0.967).
• Increased cross-scanner Gleason grading robustness by +7.4% Cohen's Kappa on PANDA compared to prior MIL models.
• Generated interpretable prototype-based attribution maps highlighting meaningful structures.

Conclusions:

• StructMIL offers a practical solution for multi-center computational pathology workflows by enhancing accuracy, interpretability, and generalization.
• The framework demonstrates significant improvements in robustness against domain shifts across scanners and institutions.
• StructMIL provides reliable and interpretable insights for cancer diagnosis and grading.