4D computed tomography super-resolution reconstruction based on tensor product and nuclear norm optimization. (January 2022)

• Record Type: Journal Article
• Authors: Zhang, Shu; Xia, Youshen
• Is Part Of: Pattern recognition. Volume 121(2022)
• Publication Date: 2022-01
• Subjects: 4D-CT -- Super-resolution -- Tensor product -- Optimization -- Nuclear norm; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108150
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A tensor product and nuclear norm optimization method for 4D-CT super-resolution. The optimal operators extract information from each dimension of LR image tensor. Efficiently training optimal operators by different kinds of tensor product. Enhancing the quality of 4D super-resolution reconstruction. Abstract: Four-dimensional computed tomography (4D-CT) has been widely used in preoperative evaluation and radiotherapy planning of lung tumors. To reduce the damage to healthy tissue, it is a better way to limit the scan time and the number of CT slices. Yet, it leads to the reduction of CT image resolution in the superior-inferior direction. To improve the resolution of the 4D-CT image, we propose a super-resolution (SR) algorithm based on tensor product and nuclear norm optimization. The proposed cost function includes a tensor fidelity term and a nuclear norm regularization term. The tensor fidelity term consists of low-resolution (LR) and high-resolution (HR) image tensors, as well as SR operators. The nuclear norm regularization term is used to preserve the operators' low-rank. The optimization problem can be effectively solved by an alternative direction method of the multipliers (ADMM) technique. The SR operators can extract useful information from each dimension of LR image tensors to enhance the equality of 4D-CT SR reconstruction. Experimental results show that the proposed method can preserve the edge details of the 4D-CT image. Moreover, quantitative comparisons show that the proposed method increases peak signal-to-noise ratio from 1.5 dB to 5.5 dB, structural similarity index from 2 % to 11 %, visual information fidelity from 6 % to 20 %, edge model-based blur metric from 5 % to 15 %, and decreases the spatial-spectral entropy-based quality index from 1 % to 5 %, compared with conventional 4D-CT SR algorithms.