Low dose CT image enhancement based on dense transformer network

Over the past year, the popular deep learning-based low-dose CT image enhancement methods mostly use the encoder-decoder structure. This structure expands the receptive field of the network by successive downsampling for deeper features. However, continuous downsampling leads to the loss of spatial information in the image, the excessively long distance between downsampling and upsampling also causes reconstruction errors. To solve the problems, this paper proposes a dense transformer network for low-dose CT image enhancement. This network uses dense connection as the basic architecture, and uses Up/Down-Tokenization in the Transformer module instead of regular up/down-sampling operations. In addition, by adding Tokons2Tokon modules between two consecutive multi-headed self-attention modules, improve the global modeling capability of the network and reduce computation. A channel fusion module is designed to replace the MLP in the standard transformer to introducing local information for reconstruction. Experiments on the AAPM LDCT dataset show that the proposed network can retain more texture details while suppressing noise compared to existing networks.