Unsupervised cycle-consistent network using restricted subspace field map for removing susceptibility artifacts...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroimaging

Background:

• Susceptibility artifacts are a significant challenge in single-shot Echo Planar Imaging (EPI), particularly in diffusion MRI.
• Existing correction methods often have limitations in accuracy and applicability across different datasets.

Purpose of the Study:

• To develop and evaluate an unsupervised deep learning model for correcting susceptibility artifacts in single-shot EPI.
• To assess the model's performance in both preclinical and clinical settings.

Main Methods:

• A novel unsupervised cycle-consistent model, DLRPG-net, was proposed, integrating deep learning (DL) with the reverse polarity-gradient (RPG) method.
• The model utilizes a DLRPG neural network to generate field maps and spin physics-based modules for image correction, employing cycle-consistency loss for training.
• Field maps are constrained to a restricted subspace of cubic splines to ensure smoothness and prevent image blurring.

Main Results:

• DLRPG-net successfully generated smooth field maps and corrected susceptibility artifacts in single-shot EPI.
• Simulated and in vivo experiments demonstrated superior performance compared to state-of-the-art artifact correction techniques.
• Ablation studies confirmed the benefits of the cycle-consistent network and restricted subspace approach.

Conclusions:

• The DLRPG-net method provides effective susceptibility artifact correction for preclinical and clinical single-shot EPI sequences.
• This deep learning approach offers a robust solution for improving image quality in diffusion MRI.