A multi-attention Uformer for low-dose CT image denoising

Keeping the number of projection views constant and reducing the radiation dose at each view is an effective way to achieve low-dose CT. This will make the reconstructed image contain high-intensity noise, which will affect subsequent image processing, analysis and diagnosis. Currently, deep learning has shown promising performance in medical image denoising. However, Transformer solely relies on a single self-attention mechanism, which fails to consider attention computation from multiple perspectives, thus limiting the performance of the model. In this paper, we propose a multi-attention coupled, U-shaped Transformer (MA-Uformer) to achieve high-performance denoising of low-dose CT images. The MA-Uformer network comprehensively uses the local information association capability of convolutional neural network (CNN) and the global information capture capability of Transformer. It integrates pixel attention mechanism, channel attention mechanism, and spatial attention mechanism to construct a coupled architecture based on multiple attention mechanisms. Compared with the four existing representative denoising networks, the network has better denoising performance and stronger ability to preserve image details.