Counterfactual Diffusion Models Provide Interpretable Explanations of Artificial Intelligence Models in Pathology.

Counterfactual diffusion models like MoPaDi enable interpretable explanations by revealing morphological features linked to AI predictions in pathology, with 4 datasets spanning multiple cancers.

• Why it matters: Understanding what features AI models rely on is crucial for trust and validation in digital pathology, as current methods offer limited insight into the image features driving predictions.
• What they did: MoPaDi combines diffusion autoencoders with task-specific classifiers to generate realistic counterfactual histopathology images, manipulating features to induce prediction shifts across various classification tasks.
• The result: The framework successfully identified morphological features such as mucinous differentiation and lymphocytic infiltration associated with model predictions, supporting hypothesis generation and model evaluation in pathology.