Abstract / Summary
Magnetic Resonance Imaging (MRI) is a cornerstone of clinical diagnostics, yet its utility is frequently hampered by motion artifacts that compromise diagnostic accuracy and necessitate costly re-scans. While deep learning has demonstrated potential for image restoration, standard architectures often fail to generalize to complex motion or risk introducing generative hallucinations. To address this, we propose the MR 2 -AttUNet, an attention-guided multi-resolution framework for retrospective motion artifact correction. The architecture integrates Convolutional Block Attention Modules (CBAM) within the encoder hierarchy, positioned between consecutive convolutional layers and down-sampling stage, enabling the model to dynamically prioritize spatial and channel-wise features to localize non-local ghosting artifacts while preserving fine parenchymal structures. Training is supported by a stochastic k-space perturbation pipeline that simulates continuous patient motion by intermittently corrupting phase-encoding lines in the frequency domain across a wide parameter distribution ( ± 5 mm and ± 7 ° ). Quantitative evaluations demonstrate that the proposed model achieves an optimal PSNR of 39.75 dB and an SSIM of 0.9854 on simulated data. When evaluated on real clinical data, the network yields a PSNR of 34.17 dB and an SSIM of 0.9448, consistently outperforming nine contemporary state-of-the-art methods, including MoCo-Net, Stacked-UNet, MC-Net and Pix2Pix, while operating at a clinical latency of 22.1 ms. Furthermore, a blinded qualitative audit of 110 real-world clinical scans by a team of board-certified consultant neurologists confirmed that the system restores diagnostic clarity without structural distortion. This optimized 2D pipeline scales to 3D stacking, enhancing diagnostic reliability while reducing repeat neuroimaging costs.