Reproduce “Conditional Diffusion Models are Medical Image Classifiers that Provide Explainability and Uncertainty for Free” on OpenResearch

About this paper

Gian Mario Favero, Parham Saremi, Emily Kaczmarek, Brennan Nichyporuk, Tal Arbel

This work explores class-conditional diffusion models for 2D medical image classification, demonstrating that they achieve competitive performance against state-of-the-art discriminative models while inherently offering explainability and robust uncertainty quantification. The approach provides crucial features for clinical trustworthiness without requiring additional architectural modifications or post-hoc analyses.

Key insights

• Class-conditional diffusion models can effectively perform medical image classification by comparing the fidelity of reconstructions conditioned on different class labels.
• A majority voting aggregation strategy for individual denoising step predictions provides more robust and accurate classification than averaging reconstruction errors.
• The generative nature of diffusion models enables direct, intrinsic explainability via counterfactual examples and offers a natural framework for quantifying predictive uncertainty without extra modules.

Summary by alphaXiv.