Evaluating Synthetic Diffusion MRI Maps created with Diffusion Denoising Probabilistic Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Generative AI models like Stable Diffusion excel at creating high-quality synthetic images.
• AI, particularly deep learning (CNNs, ViTs), is crucial for medical and neuroimaging tasks, demanding interpretability.
• Diffusion Tensor Imaging (DTI) is vital for analyzing white matter tracts in the brain.

Purpose of the Study:

• To train Latent Diffusion Models (LDM) and Denoising Diffusion Probabilistic Models (DDPM) for generating synthetic DTI maps.
• To evaluate the realism and diversity of generated synthetic DTI data.
• To assess the utility of synthetic DTI data as augmentation for improving AI classifier performance.

Main Methods:

• Trained LDMs and DDPMs on real 3D DTI scans to generate synthetic DTI maps of mean diffusivity.
• Evaluated synthetic data quality using Maximum Mean Discrepancy (MMD) and multi-scale Structural Similarity Index Measure (MS-SSIM).
• Trained a 3D Convolutional Neural Network (CNN) sex classifier using combinations of real and synthetic DTI data for augmentation.

Main Results:

• The diffusion models successfully generated realistic and diverse synthetic DTI maps.
• Quantitative metrics (MMD, MS-SSIM) confirmed the high quality of the synthetic data.
• Training a 3D CNN sex classifier with synthetic DTI data as augmentation showed potential performance improvements.

Conclusions:

• The developed diffusion models efficiently produce high-quality synthetic DTI data.
• Synthetic DTI data can serve as effective data augmentation for AI models in neuroimaging.
• This approach holds promise for advancing interpretable AI-driven diagnostics and neuroscience research.