MDST: multi-domain sparse-view CT reconstruction based on convolution and swin transformer.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Sparse-view computed tomography (SVCT) reduces radiation dose and acquisition time but faces reconstruction challenges.
• Existing Convolutional Neural Network (CNN) methods struggle to model global context dependencies in CT images.
• The locality of CNNs limits their efficiency in reconstructing complex structural information.

Purpose of the Study:

• To develop a novel deep learning network for improved SVCT image reconstruction.
• To address the limitations of CNNs in capturing global context and structural information.
• To enhance the quality of medical images reconstructed from sparse-view data.

Main Methods:

• A Multi-Domain Optimization Network based on Convolution and Swin Transformer (MDST) was developed.
• MDST utilizes Swin Transformer blocks in both projection and image domains to model global and local features.
• The network comprises initial reconstruction and residual-assisted reconstruction modules to suppress artifacts and preserve details.

Main Results:

• MDST effectively alleviates fine detail loss caused by information attenuation.
• Experiments on CT lymph node and walnut datasets demonstrate improved reconstruction quality.
• The network shows robustness in reconstructing images from projections with varying noise levels.

Conclusions:

• The MDST network demonstrates significant potential for SVCT reconstruction, outperforming CNN-based approaches.
• The use of Transformer architecture proves effective in modeling global dependencies for SVCT.
• MDST offers a robust solution for enhancing medical image quality in low-dose CT protocols.