Unsupervised Deep Learning for Susceptibility Distortion Correction in Connectome Imaging.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• High-resolution diffusion MRI (dMRI) data preprocessing often retains residual susceptibility-induced distortions.
• Existing methods like HCP-Pipeline's topup may not fully correct these distortions.

Purpose of the Study:

• To introduce an unsupervised deep learning method, DistoRtion Correction Net (DrC-Net), for residual distortion correction in dMRI.
• To improve the accuracy of dMRI data by utilizing fiber orientation distribution (FOD) images instead of B0 images.

Main Methods:

• DrC-Net employs a U-Net architecture to extract features from FOD images and estimate a deformation field.
• A transformer network is integrated to propagate deformation features and backpropagate losses.
• The model was trained on 60 subjects from the Human Connectome Project (HCP) dataset.

Main Results:

• DrC-Net demonstrated comparable performance on unseen data (40 subjects) as on the training set.
• Evaluation using mean squared difference of fractional anisotropy (FA) and minimum angular difference showed significant improvement over the topup method.
• The method effectively corrected susceptibility-induced distortions.

Conclusions:

• DrC-Net offers a robust and effective unsupervised deep learning approach for correcting residual susceptibility-induced distortions in high-resolution dMRI.
• Utilizing FOD images provides more reliable contrast information for distortion correction compared to traditional B0 methods.
• The proposed method shows significant potential for enhancing the quality of dMRI data analysis.