Generating diffusion MRI scalar maps from T1-weighted images using Reversible GANs.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Diffusion Tensor Imaging (DTI) offers critical insights into brain tissue microstructure but requires time-consuming data acquisition.
• Data scarcity and accessibility challenges limit the widespread use of DTI in clinical and research settings.
• Existing pipelines for deriving diffusion measures from structural MRI are often multi-step and complex.

Purpose of the Study:

• To investigate the generation of synthetic DTI scalar maps, specifically mean diffusivity (MD), from structural T1-weighted brain MRI.
• To assess the quality and utility of these synthetic DTI maps in downstream classification tasks.
• To evaluate the potential of a reversible generative adversarial network (RevGAN) for single-step T1 to DTI translation.

Main Methods:

• Employed a reversible generative adversarial network (RevGAN) for direct translation from T1-weighted MRI to synthetic DTI MD maps.
• Assessed the synthetic maps' utility in two classification tasks: sex classification and Alzheimer's disease classification.
• Compared the performance of machine learning models trained on real DTI maps versus RevGAN-generated synthetic DTI maps.
• Evaluated the generalization capability of models trained on synthetic data to an independent Indian cohort (NIMHANS).

Main Results:

• RevGAN successfully generated synthetic DTI MD maps from T1-weighted MRI in a single step.
• Models trained on synthetic DTI maps achieved competitive accuracy in sex and Alzheimer's disease classification compared to models trained on real DTI maps.
• The synthetic DTI maps retained meaningful microstructural information relevant for downstream analysis.
• Models demonstrated good generalization to an external dataset from a different population.

Conclusions:

• RevGAN offers a promising approach for generating synthetic DTI scalar maps, mitigating data scarcity in neuroimaging.
• Synthetic DTI data can serve as a viable alternative or supplement to real DTI data for certain analytical tasks and data augmentation.
• This single-step translation method enhances the accessibility and efficiency of diffusion-derived measure analysis in neuroimaging workflows.