A novel super-resolution CT image reconstruction via semi-supervised generative adversarial network
- PDF / 3,537,106 Bytes
- 16 Pages / 595.276 x 790.866 pts Page_size
- 47 Downloads / 186 Views
(0123456789().,-volV)(0123456789(). ,- volV)
S.I. : DEEP LEARNING APPROACHES FOR REALTIME IMAGE SUPER RESOLUTION (DLRSR)
A novel super-resolution CT image reconstruction via semi-supervised generative adversarial network Xin Jiang1 • Mingzhe Liu1 • Feixiang Zhao1 • Xianghe Liu1 • Helen Zhou2 Received: 10 April 2019 / Accepted: 6 April 2020 Ó Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract Reconstruction of super-resolution CT images using deep learning requires a large number of high-resolution images. However, high-resolution images are often limited to access due to CT performance and operation factors. In this paper, a new semi-supervised generative adversarial network is presented to accurately recover high-resolution CT images from low-resolution counterparts. We use a deep unsupervised network of 16 residual blocks to design the generator and build a discriminator based on a supervised network. We also apply a parallel 1 9 1 convolution operation to reduce the dimensionality of each hidden layer’s output. Four types of loss functions are presented to build a new one for enforcing the mappings between the generator and discriminator. The bulk specification layer in the commonly used residual network is removed to construct a new type of residual network. In terms of experiments, we conduct an objective and subjective comprehensive evaluation with several state-of-the-art methods. The comparison results show that our proposed network has better advantages in super-resolution image reconstruction. Keywords Super-resolution Computed tomography images Residual blocks Generative adversarial network
1 Introduction Super-resolution CT image reconstruction based on deep learning has been a hot topic in academic and medical areas [1–4]. There are normally three types of super& Xin Jiang [email protected] Mingzhe Liu [email protected] Feixiang Zhao [email protected] Xianghe Liu [email protected] Helen Zhou [email protected] 1
State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, Sichuan, China
2
School of Engineering, Manukau Institute of Technology, Auckland, New Zealand
resolution reconstruction methods [5]: interpolation-based [6, 7], reconstruction-based [8, 9], and learning-based [10]. The algorithms based on interpolation include bilinear interpolation, nearest-neighbor interpolation, and bicubic interpolation [11]. This kind of algorithm may lack good adaptability. The whole image looks blurred, and the edge contour is too smooth because the complexity of image features is ignored. Reconstruction-based super-resolution method includes iterative back-projection [12, 13], maximum posterior probability [14, 15], convex set projection [16], and hybrid MAP/POCS [17]. This type of method eliminates the blockiness of the interpolation algorithm to a certain extent, which improves the quality of the reconstructed image. However, when inputting a smaller size image, the effective prior knowledge is not rich enough to m
Data Loading...
