Diffusion-based generative fiber orientation restoration from severe signal loss in diffusion-weighted MRI

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) is crucial for brain connectivity analysis.
• Susceptibility-induced distortion corrupts dMRI data, impacting Fiber Orientation Distributions (FODs) reconstruction.
• Existing distortion correction methods are insufficient for certain brain regions.

Purpose of the Study:

• To develop a novel diffusion model for restoring distorted FODs.
• To address the unique challenges of applying generative models to 4D FOD data represented by spherical harmonics (SPHARM).
• To improve the accuracy of fiber tracking and connectivity analysis.

Main Methods:

• Proposed a novel diffusion model incorporating volume-order encoding for SPHARM-based FODs.
• Utilized cross-attention features across SPHARM orders to capture dependencies.
• Conditioned the model with surrounding FODs for geometric coherence.
• Trained and validated on UK Biobank (n=1315) and HCP-Aging (n=679) datasets.

Main Results:

• Demonstrated high accuracy in restoring FODs in the brainstem and orbitofrontal lobes.
• Showed improved corticospinal tract (CST) tractography performance on both datasets.
• Achieved superior CST reconstruction from single-PE direction restored FODs compared to merged two-PE data on HCP-Aging.

Conclusions:

• The proposed diffusion model effectively restores distorted FODs, enhancing dMRI analysis.
• This method offers a promising solution for improving tractography and connectivity studies, particularly in challenging brain regions.
• Generative restoration of FODs can potentially reduce the need for multi-direction data acquisition.