Purpose: To design an unsupervised deep neural model for correcting susceptibility artifacts in single-shot Echo Planar Imaging (EPI) and evaluate the model for preclinical and clinical applications. Methods:This work proposes an unsupervised cycle-consistent model based on the restricted subspace field map to take advantage of both the deep learning (DL) and the reverse polarity-gradient (RPG) method for single-shot EPI.The proposed model consists of three main components: (1) DLRPG neural network (DLRPG-net) to obtain field maps based on a pair of images acquired with reversed phase encoding; (2) spin physical model-based modules to obtain the corrected undistorted images based on the learned field map; and (3) cycle-consistency loss between the input images and back-calculated images from each cycle is explored for network training. In addition, the field maps generated by DLRPG-net belong to a restricted subspace, which is a span of predefined cubic splines to ensure the smoothness of the field maps and avoid blurring in the corrected images. This new method is trained and validated on both preclinical and clinical datasets for diffusion MRI. Results:The proposed network could effectively generate smooth field maps and correct susceptibility artifacts in single-shot EPI. Simulated and in vivo preclinical/clinical experiments demonstrated that our method outperforms the state-of-the-art susceptibility artifact correction methods. Furthermore, the ablation experiments of the cycle-consistent network and the restricted subspace in generating field maps did show the advantages of DLRPG-net. Conclusion:The proposed method (DLRPG-net) can effectively correct susceptibility artifacts for preclinical and clinical single-shot EPI sequences.

The transient severe motion may cause severe image degradation during gadoxetic acid-enhanced arterial phase imaging. This work proposes a new dual-domain unsupervised motion artifacts disentanglement network for motion correction related to gadoxetic acid-enhanced MRI. We assume that motion-free images and motion-corrupted images belong to the different domains, then the motion correction is converted to the image-to-image translation problem. The image-to-image translation within the same domain is designed to constrain autoencoders to learn the feature representation. And the cross-domain translation explores the cycle consistency in the absence of paired images. Experimental results demonstrate that our method can effectively reduce artifacts in the gadoxetic acid-enhanced images.