Reproduce “Additive MIL: Intrinsically Interpretable Multiple Instance Learning for Pathology” on OpenResearch

About this paper

Syed Ashar Javed, Dinkar Juyal, Harshith Padigela, Amaro Taylor-Weiner, Limin Yu, Aaditya Prakash

PathAI Inc. researchers developed Additive MIL, a Multiple Instance Learning framework for digital pathology that provides intrinsic, exact, and class-wise spatial credit assignment for model predictions. The approach maintains predictive performance comparable to or exceeding existing non-additive MIL models while offering clear visual explanations aligned with expert annotations.

Key insights

• Intrinsic interpretability can be achieved in MIL models for digital pathology without sacrificing predictive performance by imposing an additive constraint on the final prediction layer.
• Exact, class-wise, and excitatory/inhibitory spatial credit assignment is critical for debugging model failures and verifying that predictions are based on biologically concordant features.
• The additive constraint may act as a regularizer, potentially limiting overfitting and enabling broad applicability to existing MIL architectures, like TransMIL, through a simple modification.

Summary by alphaXiv.