Manifold-aware synthesis of high-resolution diffusion from structural imaging.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Diffusion-weighted imaging (DWI) has limited spatial resolution compared to T1w images.
• High-resolution T1w images contain valuable information for enhancing DWI detail.
• Current deep generative models struggle with the non-Euclidean nature of diffusion imaging.

Purpose of the Study:

• To develop a novel deep learning architecture for synthesizing diffusion MRI data.
• To generate diffusion tensors (DT) and diffusion orientation distribution functions (dODFs) directly from high-resolution T1w images.
• To overcome limitations of existing models in producing physically plausible diffusion images.

Main Methods:

• Proposed the first Riemannian network architecture for direct DT and dODF generation.
• Integrated the log-Euclidean Metric into the learning objective to ensure mathematical validity.
• Validated generated diffusion data by comparing tractograms with ground truth.

Main Results:

• Achieved >23% improvement in fractional anisotropy mean squared error (FA MSE).
• Improved cosine similarity between principal directions by almost 5% compared to baselines.
• Generated tractograms showed <3% difference in length and <1% difference in volume, with visually similar shapes.

Conclusions:

• The Riemannian network enables mathematically valid synthesis of diffusion MRI data from T1w images.
• The method significantly improves quantitative metrics and tractogram similarity.
• Results suggest a link between brain geometry and white matter architecture, warranting further investigation.