Enhanced Feature Contrastive Self-Supervised Learning Method Based on Deep Unfolding POCS for Accelerated Parallel MRI Reconstruction

The application of deep learning has significantly accelerated magnetic resonance imaging (MRI). However, these methods encounter substantial challenges when fully sampled datasets are unavailable. Self-supervised techniques enable network training using only undersampled MRI data. Nevertheless, the absence of fully sampled reference data often limits reconstruction quality. To address this limitation, we propose a parallel framework called the feature contrastive self-supervised learning method based on DUPOCS (FC-DUPOCS). This framework leverages deep unfolding projection onto convex sets (POCSs) and maximizes the similarities between parallel reconstruction results through contrastive learning, with the goal of improving self-supervised reconstruction quality. In FC-DUPOCS, we design deep unfolding POCS (DUPOCS) as the backbone network by unrolling the POCS-based iterative optimization process used to solve the L1SPIRiT model. Moreover, to enhance the model’s generalizability across various undersampling masks and acceleration factors (AFs), we introduce an enhanced version, EFC-DUPOCS, which simulates different undersampling masks during re-undersampling. Experimental results show that EFC-DUPOCS can effectively generalize to other unseen undersampling masks and AFs. It significantly outperforms other self-supervised methods and achieves reconstruction quality comparable to supervised methods in in-distribution (ID) settings. Additionally, FC-DUPOCS can be trained unsupervised, yielding superior reconstruction quality compared to classical scan-specific methods. Pretrained parameters on arbitrary datasets further enhance reconstruction performance. The proposed methods effectively address the limitations of self-supervised MRI reconstruction by improving reconstruction quality and model generalizability without requiring fully sampled data. These methods offer a promising alternative to supervised approaches, significantly reducing data acquisition time and cost while maintaining high reconstruction quality, enhancing the practicality of deep learning-based MRI reconstruction in clinical settings.