Integrating Anatomical Priors into a Causal Diffusion Model.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• 3D brain MRI studies often detect subtle morphometric differences difficult to see visually.
• High MRI acquisition costs limit study sizes and statistical power.
• Existing counterfactual image generation models lack anatomical precision for MRIs.

Purpose of the Study:

• To develop a novel generative framework for creating anatomically plausible counterfactual 3D brain MRIs.
• To preserve subtle, medically relevant local variations in brain structure.
• To enhance the utility of synthetic MRIs in neuroimaging research.

Main Methods:

• Proposed the Probabilistic Causal Graph Model (PCGM), integrating voxel-level anatomical constraints into a diffusion framework.
• Utilized a probabilistic graph module to capture anatomical priors, translated into spatial binary masks.
• Employed a 3D ControlNet to encode masks, constraining a counterfactual denoising UNet, and a 3D diffusion decoder for MRI generation.

Main Results:

• PCGM generated structural brain MRIs of superior quality compared to baseline methods across multiple datasets.
• For the first time, brain measurements from PCGM-generated counterfactuals replicated subtle disease effects on cortical regions.
• Demonstrated the potential of synthetic MRIs for studies on subtle morphological differences.

Conclusions:

• PCGM effectively generates high-quality, anatomically plausible 3D brain MRIs by incorporating explicit anatomical constraints.
• The method represents a significant advancement for using synthetic MRIs in research focused on subtle morphometric variations.
• PCGM's ability to replicate disease-related findings validates its utility in neuroscience research.