Development of Multi-Stage Data Consistency Integrated U-Net Model for Undersampled MRI k-Space Reconstruction

Tóm tắt

Reducing magnetic resonance imaging (MRI) acquisition time is critical to minimize patient discomfort, motion artifacts, and overall scanning costs. Compressed sensing (CS) has been widely investigated as a promising solution for MRI acceleration by reconstructing images from undersampled k-space data. However, traditional CS-based methods often suffer from instability, long reconstruction times, and limited performance due to their reliance on handcrafted regularization and iterative optimization schemes. Recent advances in deep learning have opened new possibilities for improving MRI reconstruction by learning data-driven priors directly from large datasets. Nevertheless, purely data-driven models may generate visually plausible images that lack strict consistency with the acquired measurements, potentially compromising diagnostic reliability. To address this limitation, we propose a deep learning-based MRI reconstruction framework that explicitly incorporates a data consistency (DC) layer. This physics-guided constraint enforces agreement between the network-updated k-space and the originally sampled measurements, thereby enhancing reconstruction fidelity. Experimental results demonstrate that the proposed model achieves a noticeable improvement in structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) compared with conventional reconstruction methods. These findings highlight the effectiveness of integrating learned priors with physics-based constraints, paving the way for faster MRI acquisition while preserving high diagnostic image quality in clinical practice.