Abstract

Compressed Sensing Magnetic Resonance Imaging (CS-MRI) accelerates MRI acquisition by reconstructing high-quality images from sparsely sampled k-space data. Current deep unfolding reconstruction networks usually rely on l_1-norm relaxation, suffering from insufficient sparsity and loss of high-frequency details. And traditional single-scale feature extraction fails to fully capture the detailed features of the MRI data. In this paper, we propose a multi-scale reconstruction network (SL_0-MSNet) unfolded from the Alternating Direction Method of Multipliers (ADMM) framework to achieve accelerated MRI reconstruction. By incorporating l_0-norm regularization and designing an adaptive smooth thresholding function to address the non-differentiability challenge in l_0-norm optimization. Meanwhile, a Multi-Scale Feature Extraction and Fusion (MSFAF) module is designed, which captures local textures and global anatomical features synergistically through a cascaded structure of parallel convolutional layers with different receptive fields. Comparative experiments on standard Brain and Knee datasets show that our method significantly outperforms several mainstream CS-MRI algorithms in reconstruction quality, particularly in complex scenarios, demonstrating high clinical application potential.