A Shepard wide residual network for fast MR image reconstruction on Undersampled k-space data

Fast Magnetic Resonance Imaging (MRI) reconstruction on undersampled k-space data accelerates MRI scans while maintaining image quality. This enables faster diagnostics and improves patient comfort. Existing techniques face several challenges, including image artifacts, long reconstruction times, overfitting, and limited generalization across different anatomies or scan settings. These remain significant technical hurdles in clinical and real-world applications. To address these issues, a novel technique called the Shepard Wide Residual Network (ShWideResNet) is proposed for MRI image reconstruction. Initially, the raw k-space data and undersampling mask are used to generate an aliased image. Then, anti-aliasing is performed using a Low-Pass Filter (LPF) to reduce artifacts and then, downsampling is performed. After this, sparse-based image reconstruction is carried out using a Low-Dimensional Manifold Model (LDMM). Then, MRI reconstruction is performed using the proposed ShWideResNet, which combines the Scaling Wide Residual Network (SWideRNet) and the Shepard Convolutional Neural Network (ShCNN). The outputs from the sparse-based reconstruction and the ShWideResNet-based reconstruction are then fused to obtain the final reconstructed image. The ShWideResNet achieves a Root Mean Square Error (RMSE) of 0.102, Peak Signal-to-Noise Ratio (PSNR) of 33.600 dB, a reconstruction time of 20.288 s, and Structural Similarity Index Measure (SSIM) of 0.949.