Integrating Gaussian mixture model with adjacent spatial adaptive transformer multi-stage network for magnetic resonance image reconstruction
Tong Hou, Hongqing Zhu, Jiahao Liu, Ning Chen, Jiawei Yan, Bingcang Huang, Weiping Lu, Suyi Yang, Ying Wang · Biomedical Signal Processing and Control · 2025
Accurate and fast magnetic resonance imaging (MRI) reconstruction using undersampled data is crucial for practical applications. This paper proposes a novel multi-stage network, termed IGT-Net, which integrates the Gaussian mixture model (GMM) with an adjacent spatial adaptive transformer (ASAT). Specifically, IGT-Net consists of a compressive sensing sampling initialization module (CSS-IM) and a reconstruction group. The CSS-IM is designed to equivalently mimic the sampling process and transmit MRI images of different sampling rates to the reconstruction group. The reconstruction group consists of the proposed transformer-based iterative shrinkage threshold algorithm (TrISTA) and transformer-based Gaussian mixture model (TrGMM), which reconstruct clear MRI images more efficiently. In TrISTA, the dynamic gradient descent module (DGDM) achieves fast and stable convergence of results throughout the iteration process . The proposed transformer-based proximal mapping module (TPMM) utilizes a designed transformer to fully integrate adjacent spatial features for MRI reconstruction, thereby avoiding unclear results that may arise from neglecting adjacent stage information. Meanwhile, TrGMM incorporates both GMM and ASAT to leverage maximum likelihood estimation for weight updates and artifact removal. Additionally, a plug-and-play ASAT is designed to effectively extract adjacent spatial features. Experiments conducted on two public datasets demonstrate that IGT-Net outperforms state-of-the-art methods and significantly improves the quality of reconstructed images.