Low-dose computed tomography denoising via hybrid transformer and residual dense network

Abstract Low-dose computed tomography (CT) has become an essential diagnostic tool, but it suffers from lower image quality and higher noise compared to normal-dose CT, leading to reduced diagnostic accuracy. To address this issue, we designed a hybrid network that leverages the capabilities of stochastic window transformer and residual dense network (RDN) for enhanced image denoising. The architecture of the hybrid network is a U-shaped network composed of encoding and decoding parts. The encoding part employs a stochastic window transformer to capture global features using a stochastic window strategy with gaussian shift, effectively reducing noise and preserving image details, while the decoding part utilizes an RDN to enhance image details by integrating information across residual structures and dense blocks, which facilitates feature reuse and improves gradient flow. Additionally, a dual-path feature enhancement module was incorporated into the proposed hybrid network to facilitate direct feature transfer between the encoding and decoding parts, ensuring to get a fusion feature by combining low- and high-level features from both paths. The experimental results demonstrated that our hybrid network significantly outperformed other existing denoising methods, achieving a peak signal-to-noise ratio of approximately 33.8, a structural similarity index measure of 0.92, and a root-mean-squared error of 8.354. Consequently, our method enhances diagnostic accuracy while reducing patient`s radiation doses, which provides an effective solution for low-dose CT image denoising.