Unsupervised Deep Learning for FOD-Based Susceptibility Distortion Correction in Diffusion MRI.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Susceptibility-induced distortion is a significant artifact in diffusion MRI (dMRI) analysis.
• Current methods, including those used in the Human Connectome Project (HCP), struggle with accuracy in regions like the brainstem.
• Existing traditional and learning-based approaches have limitations in precise distortion correction.

Purpose of the Study:

• To introduce a novel deep learning framework, DistoRtion Correction Net (DrC-Net), for accurate susceptibility-induced distortion correction in dMRI.
• To leverage fiber orientation distribution (FOD) data for improved distortion correction.
• To enhance the accuracy of dMRI analysis, particularly in challenging anatomical areas.

Main Methods:

• Developed a deep learning framework (DrC-Net) integrating a U-Net and a spatial transformer network.
• Utilized 4D fiber orientation distribution (FOD) images derived from dMRI.
• Trained and validated the framework on Human Connectome Project (HCP) and Human Connectome Low Vision (HCLV) datasets.

Main Results:

• DrC-Net demonstrated significant improvements over traditional (topup, FODReg) and deep learning (S-Net, flow-net) methods.
• Achieved better accuracy in terms of mean squared difference (MSD) of fractional anisotropy (FA) images and minimum angular difference.
• Showed superior performance in both white matter and brainstem regions.
• DrC-Net provides rapid displacement field prediction (seconds), outperforming FODReg in speed.

Conclusions:

• The proposed DrC-Net framework offers a highly accurate and efficient solution for susceptibility-induced distortion correction in dMRI.
• This deep learning approach significantly enhances dMRI data quality, especially in anatomically complex regions.
• DrC-Net represents a substantial advancement in dMRI analysis, offering both improved accuracy and speed.