Medical Images Super Resolution Reconstruction based on Residual Network
Dongxu Zhao, Fang Zhang, Wen Wang, Zhitao Xiao, Lei Geng · 2021
Low-quality medical ultrasound images have serious spots, noise, and weak borders between similar tissues, which affect the accuracy of segmentation of human organs and lesions in the image to a certain extent. These problems seriously hinder the subjective diagnosis of doctors and intelligent healthcare assisted diagnosis. In this paper, a super-resolution approach is proposed based on deep learning that aims to upscale the prostate transrectal ultrasound (TRUS) image. The proposed method is based on the TRUS dataset, which is termed ultrasound image super-resolution (USSR). Firstly, this method comprises a deep residual network which through extracting feature to learns a complex mapping between low and high projections. Secondly, skip connection is introduced to enhance shallow information transmission. Thirdly, some typical ideas and sub-pixel convolution are adopted in the upsample part in order to ensure reconstruction speed. Experiments demonstrate that the USSR method proposed in this paper can reconstruct high-resolution medical images of different scales effectively, and reconstruct the internal texture and edge information of medical images accurately. Both the subjective qualitative results and the objective quantitative evaluation results show that the USSR method gives good reconstruction performance.