DreaMR: Diffusion-Driven Counterfactual Explanation for Functional MRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Deep learning models excel at detecting cognition-related variables from functional MRI (fMRI) data.
• Interpreting these complex models and their association with specific brain regions remains a significant challenge.
• Existing explanation methods like attribution and perturbation have limitations in sensitivity and specificity.

Purpose of the Study:

• To introduce DreaMR, the first diffusion-driven counterfactual method for high-fidelity fMRI interpretation.
• To address limitations of existing counterfactual generation methods in sample fidelity.
• To improve the interpretability of deep learning models applied to neuroimaging data.

Main Methods:

• DreaMR utilizes diffusion-based resampling of fMRI data to generate counterfactual samples.
• It computes the difference between original and counterfactual samples for model explanation.
• Employs a fractional multi-phase-distilled diffusion prior and a transformer architecture for efficiency and spatiotemporal context.

Main Results:

• DreaMR demonstrates superior sample generation fidelity compared to state-of-the-art counterfactual methods.
• Achieves improved inference efficiency without compromising data fidelity.
• Effectively accounts for long-range spatiotemporal dependencies in fMRI scans.

Conclusions:

• DreaMR offers a significant advancement in explaining deep learning models for fMRI.
• Provides a high-fidelity and efficient approach for neuroimaging data interpretation.
• Enables more reliable association of brain regions with cognitive variables.