Diffusion Model-based FOD Restoration from High Distortion in dMRI.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) data is commonly modeled using Fiber Orientation Distributions (FODs).
• Susceptibility-induced distortions in dMRI corrupt FOD reconstruction, hindering tractography and connectivity analysis, particularly in brain regions like the brain stem.
• Existing generative models, like diffusion models, show promise for image restoration but face challenges with the 4D nature and order-dependency of FODs.

Purpose of the Study:

• To develop a novel diffusion model for restoring Fiber Orientation Distributions (FODs) affected by imaging artifacts.
• To address the unique challenges of applying diffusion models to 4D FOD data, including spherical harmonics (SPHARM) order dependency.
• To improve the accuracy of FOD reconstruction and enhance tractography performance in distorted brain regions.

Main Methods:

• Proposed a novel diffusion model incorporating volume-order encoding to generate FOD volumes across all SPHARM orders.
• Integrated cross-attention features across SPHARM orders to capture inter-order dependencies during FOD generation.
• Conditioned the model on surrounding low-distortion FODs to preserve geometric coherence in high-distortion areas.
• Trained and validated the model using UK Biobank data (n=1315), with ground truth testing on n=43.

Main Results:

• Demonstrated high accuracy in generated FOD volumes and FOD peaks using root mean square errors and angular errors on a ground truth test set.
• Successfully applied the method to a large dataset (n=1172) with significant brain stem distortion, restoring FOD integrity.
• Showcased substantial improvements in tractography performance in brain regions affected by distortion artifacts.

Conclusions:

• The proposed diffusion model effectively restores signal loss in FODs caused by distortion artifacts.
• The model's architecture, including volume-order encoding and cross-attention, successfully handles the complexity of 4D FOD data.
• This restoration technique significantly enhances the reliability of fiber tracking and connectivity analysis in challenging neuroimaging datasets.