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

Area of Science:

• Artificial Intelligence
• Neuroimaging
• Medical Diagnostics

Background:

• Deep learning models like CNNs and ViTs are used for neuroimaging tasks, but often require large datasets and lack interpretability.
• Generative AI, particularly denoising diffusion models, can generate high-quality synthetic data and learn complex data distributions.
• Extending generative models to medical neuroimaging offers potential for enhanced diagnostics and research.

Approach:

• Trained conditional latent diffusion models (LDM) and denoising diffusion probabilistic models (DDPM) on 3D T1-weighted MRI scans, conditioning on Alzheimer's disease (AD) diagnosis.
• Investigated methods to overcome limitations in dataset size, compute time, and memory, testing various model sizes, pretraining effects, and training durations.
• Evaluated synthetic MRI quality using realism and diversity metrics, and assessed diffusion models' ability to conditionally sample MRI brains for disease classification.

Key Points:

• Diffusion models generated synthetic MRI data that improved an AD classifier's performance by over 3% using only 500 real training scans.
• Implicit classifier-free guidance was used to generate counterfactual healthy brain scans from individual patient data, preserving subject-specific details.
• Personalized disease maps were created from counterfactual images to highlight potential AD effects on an individual's brain anatomy.

Conclusions:

• The diffusion model approach efficiently generates realistic and diverse synthetic neuroimaging data.
• This method shows promise for creating interpretable AI-based maps for neuroscience research and clinical AD diagnostic applications.
• Conditional generative models offer a powerful tool for augmenting limited medical imaging datasets and advancing personalized medicine.