Self-Supervised Learning Framework With Under-Balanced Loss Optimization for Point of Care MRI Image Reconstruction in 6G-Driven Edge Networks

Abstract

Self-supervised learning frameworks in the 6G-driven edge networks provide powerful instant MRI image diagnostic capabilities for the process of point of care. Although many deep learning self-supervised frameworks are used to train-related models to improve magnetic resonance imaging (MRI) image reconstruction, there is still room for improvement in model training convergence acceleration and MRI image reconstruction quality. To address the above issues, first, this article proposes a self-supervised learning framework, which combines the real-time computing power of the edge network driven by 6G networks to accelerate the training convergence of the MRI image reconstruction model and improve the quality of the reconstructed image. Second, the proposed framework innovatively introduces an under-balanced loss optimization structure and applies heterogeneous loss functions at different positions of the model. Finally, this article proposes AttentionFISTA-Net, which integrates the convolutional attention module into FISTA-Net to enhance the MRI image reconstruction effect. Experimental results on the IXI dataset compared with the traditional self-supervised network show that the proposed model performs better in the under-sampled dataset with acceleration rates of 4 and 8, respectively. The peak signal-to-noise ratio (PSNR) metric improves $0.021$ and $0.61$ respectively, and the structure similarity index measure (SSIM) metric improves $0.5*{10}^{-3}$ and $8.2*{10}^{-3}$, respectively.