Aligning Cellular Sheaves with Classifier Attention for Interpretable Weakly-Supervised Pathology Localization

Researchers propose a novel approach combining cellular sheaves with attention-based multiple instance learning to improve interpretability in weakly-supervised pathology image classification. The method achieves 0.940 patch-level AUC on Camelyon16 and successfully aligns attention maps with diagnostic regions, addressing a critical gap where models classify correctly without focusing on actual lesions.

This research tackles a fundamental problem in medical AI deployment: the reliability gap between model accuracy and model interpretability. While attention-based multiple instance learning has achieved near-saturation performance on Camelyon16 slide-level classification, the corresponding attention maps often fail to highlight actual pathological features, creating clinical trust issues. The integration of cellular sheaves—a mathematical framework for detecting local disagreement in graph-structured data—represents a principled approach to this verification problem.

The innovation centers on attention-conditional consistency, which leverages the classifier's attention weights to determine which neighboring patches should exhibit feature agreement. This constrains the sheaf disagreement field to track diagnostic content rather than texture variations, fundamentally improving localization quality. The ablation study demonstrating that joint training produces gains unattributable to the loss function alone validates that the approach requires co-adaptation of both the classifier and sheaf components.

For clinical deployment, this dual-map output provides complementary explanations for predictions, enhancing interpretability crucial for regulatory compliance and clinician confidence. The successful transfer to Camelyon17 without retraining indicates the learned representations capture generalizable diagnostic patterns. The architecture bridges a critical gap between performance metrics and clinical usability—areas where many AI systems fail in real-world healthcare settings.

The research demonstrates how mathematical frameworks from topology can address practical machine learning interpretability challenges. Future applications may extend this approach to other vision tasks requiring both high accuracy and explainability, particularly in regulated industries where model decisions must be auditable and trustworthy.

• →Cellular sheaves combined with attention-based learning achieve 0.940 patch-level AUC for tumor localization on histopathology images
• →Attention-conditional consistency ensures model explanations align with actual diagnostic features rather than spurious patterns
• →Joint training of classifier and sheaf components proves essential; frozen baselines show minimal improvement, confirming synergistic gains
• →Trained models successfully transfer to new datasets without retraining while maintaining high localization and attention performance
• →Dual attention and disagreement maps provide clinicians with complementary explanations, addressing a critical trust gap in medical AI deployment