DCIPSR: A Self-supervised Networks with Signal Regularization for Multi-Contrast MRI Reconstruction

Motivation: Limitations on scan times continue to hinder the integration of Multi-contrast magnetic resonance imaging protocols into standard care practices. Meanwhile, supervised methods are limited by high-quality ground-truth training data and restricted generalization ability. Goal(s): To reconstruct undersampled MCMRI with high generalization ability and leading performance. Approach: We proposed a self-supervised MCMRI reconstruction method with Densely Connected Image Prior(DCIP) and Signal Regulation(SR). Results: Compared with conventional and unsupervised deep-learning algorithms, DCIPSR well removed the aliasing artifacts and achieved the leading performance. Impact: We proposed a self-supervised MCMRI reconstruction method with Densely Connected Image Prior(DCIP) and Signal Regulation(SR). The success of DCIPSR under high undersampled rate indicates the potential to reconstruct MCMRI when large training dataset is unavailable.