TTMRI: Multislice texture transformer network for undersampled MRI reconstruction

• DOI: 10.1049/ipr2.13089
• Publication Date: 2024-03-27T09:58:45Z
• Authors: Xiaozhi Zhang; Liu Zhou; Yaping Wan; Bingo Wing‐Kuen Ling; Dongping Xiong
• Keywords: Texture (cosmology)

Abstract Magnetic resonance imaging (MRI) is a non‐interposition technique that provides rich anatomical and physiological information. Yet it limited by the long time. Recently, deep neural networks have shown potential to significantly accelerate MRI. However, most of these approaches ignore correlation between adjacent slices in MRI image sequences. In addition, existing learning‐based methods for are mainly based on convolutional (CNNs). They fail capture long‐distance dependencies due small receptive field. Inspired feature similarity impressive performance Transformer exploiting dependencies, novel multislice texture transformer network presented undersampled reconstruction (TTMRI). Specifically, proposed TTMRI consisted four modules, namely extraction, calculation, transfer synthesis. It takes three as inputs, which middle one target be reconstructed, other two auxiliary images. The multiscale features extracted extraction module their inter‐dependencies calculated calculation module, respectively. Then relevant transferred fused synthesis module. By considering inter‐slice correlations leveraging architecture, joint learning across encouraged. Moreover, can stacked with multiple layers recover more information at different levels. Extensive experiments demonstrate outperforms state‐of‐the‐art both quantitative qualitative evaluationsions.