Magnetic resonance image reconstruction based on two-stage progressive network

In clinical applications, medical image reconstruction is crucial for extracting complementary information and restoring image quality. However, existing deep learning-based reconstruction methods suffer from the following two problems: first, they ignore the frequency-domain information specific to magnetic resonance imaging, and second, most of the traditional networks are single-end-to-end networks, which lack effective attention to feature information. We suggest a unique two-stage progressive network (TSPNet) to solve these issues. Our approach employs two lightweight Unets as subnetworks for progressive recovery, and the network parts are carefully designed. During the feature extraction phase, we creatively create a Dual Domain Feature Extraction Module (DDFEM) to effectively extract rich features. Subsequently, we create a hybrid attention module (HAM) to effectively process and integrate the retrieved data. Experimental results show that our model outperforms other state-of-the-art methods and reconstructs images with stronger texture information and superior visual quality.