Counterfactual Diffusion Models for Interpretable Morphology-based Explanations of Artificial Intelligence Models in...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Biomarker discovery

Background:

• Deep learning models can identify predictive and prognostic biomarkers from histopathology whole slide images.
• A significant challenge in applying deep learning to histopathology is the lack of interpretability of these models.

Purpose of the Study:

• To develop and validate a novel method for generating counterfactual mechanistic explanations in histopathology.
• To enhance the interpretability of deep learning models used in digital pathology by revealing key morphological drivers of predictions.

Main Methods:

• Developed MoPaDi (Morphing histoPathology Diffusion), a system utilizing diffusion autoencoders to manipulate pathology image patches.
• MoPaDi alters image morphology to flip biomarker status, incorporating multiple instance learning for weakly supervised tasks.
• Validated on four datasets for tissue/cancer type classification, slide origin, and microsatellite instability biomarker prediction.

Main Results:

• MoPaDi demonstrated excellent image reconstruction quality (MS-SSIM 0.966-0.992) and strong classification performance (AUCs 0.76-0.98).
• Counterfactual images generated for tissue-type classification were highly realistic, with 63.3-73.3% correctly identified in a user study.
• Pathologists identified meaningful morphological features from counterfactual images across various classification tasks.

Conclusions:

• MoPaDi successfully generates realistic counterfactual explanations for deep learning predictions in histopathology.
• The method effectively reveals critical morphological features that influence model predictions, thereby improving model interpretability.
• This approach holds promise for advancing biomarker discovery and understanding in digital pathology.