Abstract Background Deep learning (DL)‐based magnetic resonance imaging (MRI) reconstruction methods have been widely applied to accelerate data acquisition. However, most of these methods rely heavily on large‐scale fully sampled labeled data for supervision, which is often difficult to obtain in clinical practice, thereby limiting their practical application. Purpose To address the aforementioned limitation, this study proposes a self‐supervised Siamese cooperative network (S 3 CNet) for high‐quality MRI reconstruction using highly undersampled k ‐space data. The goal is to reduce reliance on fully sampled data while maintaining excellent reconstruction performance. Methods A unified two‐stage self‐supervised reconstruction framework was designed, which is compatible with different sampling patterns and network architectures. In the first stage, undersampled k ‐space data were randomly split into two disjoint subsets. By introducing measurement consistency loss to align reconstructed k ‐space with measured data and cross‐consistency loss to suppress noise via Noise2Noise‐inspired supervision, the framework effectively suppresses noise propagation from measured data to reconstructed results. In the second stage, the denoised outputs of the first stage were used as pseudo‐labels, and reconstruction consistency loss was adopted to constrain the sampling‐reconstruction pipeline, which enhances the network's generalization ability to unseen data. These three losses constitute a hybrid self‐supervised loss function, which enables effective training of the Siamese network without the need for fully sampled ground truth. Results Experimental results on the T1‐ and T2‐weighted brain FastMRI dataset showed that S 3 CNet outperformed state‐of‐the‐art self‐supervised methods under various undersampling rates and sampling patterns. Notably, S 3 CNet's performance was comparable to that of fully supervised methods. Conclusions S 3 CNet effectively overcomes the critical dependency of DL‐based MRI reconstruction on fully sampled training data. It achieves both high reconstruction quality and strong generalization, providing a practical solution for accelerated clinical MRI.
Geng et al. (Sun,) studied this question.