Abstract

Info

Author(s):: XIE Fengxue1; YANG Fan1; FENG Wei1; ZENG Leilei2; MIAO Yuehong1; LEI Pinggui3; 1. School of Biology and Engineering, Guizhou Medical University, Guiyang 550025, China 2. School of Big Health, Guizhou Medical University, Guiyang 550025, China 3. Department of Radiology, the Affiliated Hospital of Guizhou Medical University, Guiyang 550004, China

Keywords:: Keywords: low-dose computed tomography image denoising deep learning multiscale parallel residual U-net

Abstract:: Objective To propose a deep learning-based method for low-dose computed tomography (LDCT) image denoising. Methods After reconstruction by filtered back projection, a deep learning model of multiscale parallel residual U-net (MPR U-net) was used for denoising the reconstructed LDCT images. The medical CT datasets of LoDoPaB-CT Challenge were used in the experiment, including 35 820 images in training set, 3 522 images in validation set and 3 553 images in test set. The denoising effect of the model was evaluated by peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Results The PSNR before and after LDCT image denoising was 28.80 and 38.22 dB, respectively, and the SSIM was 0.786 and 0.966, respectively. The average processing time was 0.03 s. Conclusion The proposed MPR U-net deep learning model can remove LDCT image noise better, improve PSNR and retain more image details.