Explaining Deep fMRI Classifiers with Diffusion-Driven Counterfactual Generation

Motivation: Deep-learning classifiers for functional MRI (fMRI) offer state-of-the-art performance in detection of cognitive states from BOLD responses, but their black-box nature hinders interpretation of results. Goal(s): Our goal was to devise a reliable method to infer the important BOLD-response attributes that drive the decisions of deep fMRI classifiers. Approach: We introduced a novel counterfactual explanation method (DreaMR) based on a new fractional, distilled diffusion prior for efficient generation of high-fidelity counterfactual samples. Results: DreaMR generated more specific and plausible explanations of deep fMRI classifiers trained for resting-state and task-based fMRI analysis than previous state-of-the-art explanation methods. Impact: The improvement in sensitivity, plausability and efficiency in explanation of deep classifiers through DreaMR may facilitate adoption of AI-based analyses in fMRI studies, thereby benefiting assessment of cognitive processes in both normal and neurological disease states.