Sparse Reconstruction of Computed Tomography Images with Transformer Enhanced U-net

An effective way to achieve low-dose computed tomography (CT) is to reduce the projection angle while maintaining the same radiation dose at each angle. However, a fewer projection angle can result in severe strip artifacts, reducing the practicality and clinical value of the image. To address this issue, a U-shaped network (Transformer Enhanced U-net, TE-unet) coupled with convolutional neural network (CNN) and multiple attention mechanisms was proposed. Firstly, a U-shaped architecture was adopted to fuse multi-scale feature information; Secondly, a module that includes CNN and multiple types of attention was proposed to extract image features; Finally, transformer blocks were added at skip connections to filter information, suppress irrelevant features, and highlight important features. This network combines the local feature extraction ability of CNN and the global information capture ability of Transformer, supplemented by various attention mechanisms, to achieve good ability to remove stripe artifacts. At 60 projection angles, compared to the classic uformer network, peak signal to noise ratio (PSNR) is 0.3178 dB higher, Structural Similarity (SSIM) is 0.002 higher, and Root Mean Square Error (RMSE) is 0.0005 lower. The experimental results show that the proposed TE-unet network reconstructs images with higher accuracy, preserves better image details, and can better suppress strip artifacts.