Attention-Enhanced Convolutional Iterative Unrolling Network for Under-Sampled MRI Reconstruction

Compressed Sensing Magnetic Resonance Imaging (CS-MRI) enables high-quality image reconstruction from a small set of the k-space data, thereby accelerating the efficiency of MRI. Existing deep unrolling-based CS-MRI algorithms predominantly utilize Convolutional Neural Network (CNN) architectures with fixed-size convolutional kernels to extract single-scale features from k-space data, resulting in the loss or blurring of important information in the reconstructed magnetic resonance (MR) image. Moreover, due to the inherent local feature dependency within CNN architectures, these methods may possess inherent limitations in capturing the global structural information present in MR data. To address the aforementioned challenges, we propose an Attention-Enhanced Convolutional Iterative Unrolling Network for CS-MRI reconstruction, called ACIU-Net, which can obtain high-quality MR reconstruction images. Specifically, we design a Multi-Scale Convolutional Feature Extractor (MFE) module to adequately capture multi-scale features of k-space data, and use convolutional blocks with residual concatenation to learn the nonlinear transformation process of the Proximal Gradient Descent (PGD) algorithm. Meanwhile, we introduce the customized Global Feature Aggregation Attention (GFA) modules within the MFE to strengthen the feature representation capability, which can adjust the importance of each sub-feature by generating distinct attention factors of the MR data, thereby enabling the capture of its global characteristics. Experimental results on the Brain, Cardiac and FastMRI datasets show that our proposed ACIU-Net significantly outperforms the current state-of-the-art CS-MRI methods.