Cellular sheaves boost pathology AI localization accuracy to 0.94 AUC

Whole-slide image classification with attention-based multiple instance learning (ABMIL) has reached near-saturation on the Camelyon16 benchmark, but its attention maps often fail to highlight the actual lesion, undermining clinical trust. A new paper from researchers Lalwani, Bhat, and Shah tackles this by introducing cellular sheaves—a graph-based mathematical structure that assigns vector spaces to each vertex and edge with consistent linear maps. Applied to weakly-supervised tumor localization, the sheaf produces a disagreement field that flags local inconsistencies between patches. The key innovation is attention-conditional consistency: the system uses the classifier's attention weights to decide which neighboring patches should agree, rather than assuming all similar features are diagnostic. Joint training of the classifier and sheaf under this objective yields a disagreement field with patch-level AUC 0.940 on Camelyon16, while the attention head jumps from its ABMIL-alone level of 0.717 to 0.953.

Ablation experiments confirm that the gains come from co-adaptation under both objectives—freezing the classifier at ABMIL values only raises the disagreement field to 0.727 and leaves attention at 0.717. The trained model transfers without any retraining to annotated slides from Camelyon17, maintaining delta AUC 0.932 ± 0.083 and attention AUC 0.955 ± 0.099. The final output provides clinicians with two correlated explanation maps: the standard attention map and a sheaf-disagreement map that both fire on the same diagnostic regions. This dual-explanation approach builds interpretability into the model itself, closing the trust gap between high slide-level accuracy and meaningful local explanations—a critical step for deploying AI in pathology workflows.