Counterfactual MRI Generation with Denoising Diffusion Models for Interpretable Alzheimer's Disease Effect Detection

Area of Science:

• Artificial Intelligence
• Neuroimaging
• Medical Diagnostics

Background:

• Generative AI, particularly denoising diffusion models, can produce high-quality synthetic images and learn complex data distributions.
• Understanding Alzheimer's disease (AD) effects on brain anatomy requires advanced analytical tools.
• Limitations in training dataset size and computational resources hinder the development of AI models for medical imaging.

Purpose of the Study:

• To train conditional latent diffusion models (LDM) and denoising diffusion probabilistic models (DDPM) for generating synthetic 3D T1-weighted MRI scans.
• To investigate the impact of Alzheimer's disease on brain anatomy at an individual level using generative AI.
• To overcome challenges in training dataset size, compute time, and memory resources for diffusion model development.

Main Methods:

• Trained conditional LDMs and DDPMs on real 3D T1-weighted MRI scans, conditioning the generation on clinical diagnosis.
• Evaluated synthetic MRI quality using metrics for realism and diversity.
• Assessed the ability of diffusion models to conditionally sample MRI brains for a 3D CNN-based disease classifier.
• Employed classifier-free guidance to generate counterfactual healthy brain scans from individual patient data.
• Generated personalized disease maps from counterfactual images to identify potential disease effects.

Main Results:

• Diffusion models generated realistic and diverse synthetic MRI data.
• Synthetic data improved the performance of an Alzheimer's disease classifier by over 3% when trained on limited real MRI scans.
• Classifier-free guidance enabled the generation of personalized counterfactual healthy brain scans, preserving subject-specific details.
• Personalized disease maps were successfully generated to highlight potential AD-related brain changes.

Conclusions:

• Conditional diffusion models can efficiently generate high-quality synthetic brain MRI data.
• The generated synthetic data effectively enhances the training of diagnostic AI models for Alzheimer's disease.
• The approach offers a novel method for creating interpretable, AI-based personalized disease maps for neuroscience research and clinical applications.