Automated Mapping of Residual Distortion Severity in Diffusion MRI.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Susceptibility-induced distortion is a common artifact in diffusion MRI (dMRI), impacting connectivity analysis.
• Existing distortion correction methods often leave residual artifacts that vary across subjects and brain regions.
• A voxel-level map of residual distortion severity is needed to guide downstream analyses.

Purpose of the Study:

• To develop and validate a supervised deep learning network for predicting residual distortion severity maps in dMRI.
• To provide a tool that can inform connectivity analyses by quantifying remaining distortions.
• To create an efficient method applicable to large-scale datasets with varying data acquisition parameters.

Main Methods:

• A supervised deep learning network was trained using the structural similarity index measure (SSIM) of fiber orientation distribution (FOD) from opposite phase encoding (PE) directions.
• The model requires only b0 images and outputs from distortion correction methods as input during testing.
• The model was trained on the HCP-Aging dataset and tested on UK Biobank data.

Main Results:

• The deep learning model demonstrated low training, validation, and test errors.
• Generated severity maps showed excellent correlation with fiber orientation distribution (FOD) integrity measures in both datasets.
• The method is highly efficient, generating a severity map in approximately one second per subject.

Conclusions:

• The proposed deep learning approach effectively generates voxel-level residual distortion severity maps for dMRI.
• This method offers a valuable tool for enhancing the accuracy of diffusion MRI connectivity analyses.
• The approach is efficient and suitable for large-scale neuroimaging studies, including those with single-direction dMRI data.