Self-Supervised Learning for Three-Dimensional Magnetic Resonance Imaging Reconstruction with Spatial Depth Attention

Reconstructing images from undersampled k-space data is crucial for accelerating MRI acquisition. While deep learning methods have shown advantages, they typically rely on fully sampled reference data, which are difficult to obtain and time-consuming. To address this issue, we propose a self-supervised method based on two parallel reconstruction networks. These parallel networks restore information from random subsets of the undersampled k-space data, while a spatial-depth attention mechanism, specifically designed for three-dimensional data, enhances feature interaction and fusion. A reconstruction loss defined in k-space helps recover frequency information, and a differential loss ensures consistency between the networks. On BraTS/IXI, our self-supervised method achieves PSNR/SSIM on par with supervised methods at 4×/8× acceleration.