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

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in histopathology

Background:

• Deep learning models excel at extracting biomarkers from histopathology images.
• Current explainable AI methods (e.g., heatmaps) offer limited insight into image features driving predictions.
• A need exists for advanced methods to understand model decision-making in digital pathology.

Purpose of the Study:

• To develop MoPaDi (Morphing histoPathology Diffusion), a framework for generating counterfactual explanations in histopathology.
• To identify morphological or stain-related features associated with deep learning model predictions.
• To enhance the interpretability of AI models in computational pathology.

Main Methods:

• MoPaDi combines diffusion autoencoders with multiple instance learning classifiers.
• The framework manipulates histopathology images to induce prediction shifts by altering classifier-associated features.
• Evaluated on diverse cancer datasets (colorectal, breast, liver, lung) for various classification tasks.

Main Results:

• MoPaDi generated perceptually realistic counterfactual images, aiding pathologists in identifying key features.
• The framework successfully highlighted morphological features (e.g., mucinous differentiation, glandular architecture, lymphocytic infiltration) linked to microsatellite instability predictions.
• Analyses indicated that morphology alterations, rather than stain variations, predominantly influenced prediction changes.

Conclusions:

• MoPaDi provides a practical approach for counterfactual explanations in computational pathology.
• The framework supports the evaluation of model-specific decision cues and facilitates hypothesis generation for biomarker discovery.
• MoPaDi enhances the interpretability of deep learning models in analyzing histopathology whole-slide images.