Abstract

Info

Author(s):: CHEN Hongchi1; 2; 3; ZHAI Ying1; 2; 3; LI Qiuxia1; 2; 3; LI Fangzuo1; 2; 3; 1. School of Medical and Information Engineering, Gannan Medical University, Ganzhou 341000, China 2.Key Laboratory of Prevention and Treatment of Cardiovascular and Cerebrovascular Diseases, Ministry of Education, Gannan Medical University, Ganzhou 341000, China 3. Jiangxi Provincial Key Laboratory of Tissue Engineering (2024SSY06291), Ganzhou 341000, China

Keywords:: Keywords: low-dose CT deep learning perceptual loss image denoising artifact reduction

Abstract:: Abstract: Objective Low-dose CT (LDCT) images significantly affect clinical diagnosis due to the substantial presence of noise and artifacts. To address the challenges such as over-smoothing of images, loss of texture details, and the presence of residual noise artifacts, a deconvolution-convolution residual network integrating multi-scale dilated convolution (MDC-DCRN) is proposed for LDCT denoising. Methods The network employed deconvolution-convolution architecture to better preserve image details and integrated an MDC module to enhance the feature extraction capabilities at different scales. Moreover, the issue of excessive image smoothing was effectively mitigated by the composite loss function combining L1 loss and perceptual loss. Results The experimental results on the Mayo dataset demonstrated that MDC-DCRN outperformed 4 classic denoising methods, namely RED-CNN, EDCNN, WGAN-RAM, and CTformer. MDC-DCRN effectively eliminated noise and artifacts while recovering more texture detail information. Compared with LDCT images, the images processed by MDC-DCRN had an average increase of 13.64% in peak signal-to-noise ratio, an average increase of 4.57% in structural similarity index, and an average decrease of 37.40% in root mean square error. Conclusion The proposed MDC-DCRN model can effectively preserve details while reducing noise from low-dose scanning, offering a novel approach to clinical LDCT image denoising.