Diffusion Models for Neuroimaging Data Augmentation: Assessing Realism and Clinical Relevance.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning in medical imaging is hindered by data scarcity, especially for rare neurodegenerative diseases.
• Generating realistic synthetic medical images is crucial for overcoming data limitations.

Purpose of the Study:

• To investigate the use of denoising diffusion probabilistic models (DDPMs) for generating synthetic 3D T1-weighted brain MRI images.
• To address challenges of limited training data and structural fidelity in medical image synthesis.

Main Methods:

• A generative pipeline using DDPMs was trained on a multicenter dataset of healthy subjects.
• Quantitative evaluation used Maximum Mean Discrepancy (MMD) to compare real and synthetic data distributions.
• Visual assessments were performed to evaluate the preservation of global and local brain structures.

Main Results:

• The DDPMs demonstrated potential in producing anatomically coherent synthetic brain MRI scans with realistic variability.
• Quantitative analysis confirmed the similarity between real and generated data distributions.
• Visual assessments indicated successful preservation of global and local brain structures in synthetic images.

Conclusions:

• DDPMs show promise for augmenting neuroimaging datasets, particularly for rare diseases, by generating synthetic 3D T1-weighted brain MRI scans.
• Synthetic data can support downstream tasks like classification and segmentation in medical imaging.
• Further research is needed to improve resolution and adapt models for specific rare disease imaging challenges.