Multi-modality conditioned variational U-net for field-of-view extension in brain diffusion MRI

Area of Science:

• Neuroimaging
• Medical Physics
• Computational Neuroscience

Background:

• Incomplete field-of-view (FOV) in diffusion magnetic resonance imaging (dMRI) hinders whole-brain white matter connectivity analysis.
• Existing deep generative models for dMRI imputation do not fully leverage multi-modality data for enhanced quality.

Purpose of the Study:

• To develop a novel framework for imputing incomplete dMRI scans by integrating diffusion features with anatomical structures.
• To evaluate the framework's ability to improve imputation quality and downstream tractography accuracy.

Main Methods:

• A new framework was proposed to impute missing dMRI data by integrating learned diffusion features from the acquired FOV with complete brain anatomical structures.
• The framework was tested on two cohorts (96 subjects) and compared against a baseline method treating T1w and dMRI data equally.

Main Results:

• The proposed framework significantly improved imputation performance, evidenced by a higher angular correlation coefficient (p < 1E-5).
• Downstream tractography accuracy was also enhanced, shown by a significant improvement in Dice score (p < 0.01).

Conclusions:

• The framework effectively utilizes multi-modality data for superior dMRI imputation compared to baseline methods.
• Improved dMRI imputation enhances whole-brain tractography, reducing uncertainty in analyzing neurodegenerative disease-associated white matter bundles.