DDCN: A Dual Deep Contrastive Network for Self-supervised MRI Reconstruction

Magnetic resonance imaging (MRI), renowned for its superior soft tissue delineation and absence of ionizing radiation, is extensively employed in clinical diagnostics and medical research. However, limited by long data acquisition procedures, MRI is often criticized by its high cost and discomfort. Recently, the integration of deep learning techniques into the realm of swift MRI reconstruction has yielded significant advancements. Nonetheless, the potential of deep learning for accelerating MRI is constrained by the challenges and substantial expense involved in obtaining labeled MR datasets. This study introduces a self-supervised dual-domain parallel framework for MRI reconstruction that facilitates fast MR imaging without fully-sampled images. Specifically, we creates two distinct subsets from undersampled MR data. This enables the model to perform rapid MR image reconstruction without fully-sampled MR images. Furthermore, feature learning and reconstruction are concurrently conducted in both the image and k-space domains. In addition, we use a siamese network to share parameters. We assess the efficacy of the proposed method using the public IXI brain dataset. The results demonstrate that our model can train models with undersampled inputs and match the quality of supervised methods at various acceleration rates.