Artifact removal for unpaired thorax CBCT images using a feature fusion residual network and contextual loss

• Published in: Journal of applied clinical medical physics Vol. 24; no. 7; pp. e13968 - n/a
• Main Authors: Zhuang, Wenqin, Li, Zheng, Liu, Haochen, Ying, Hu, Yan, Mo
• Format: Journal Article
• Language: English
• Published: United States John Wiley & Sons, Inc 01-07-2023; John Wiley and Sons Inc
• Subjects: Accuracy; Algorithms; artifacts correction; Cone-Beam Computed Tomography - methods; contextual loss; Deep learning; FFRN; Humans; Image Processing, Computer-Assisted - methods; Methods; Neural networks; Phantoms, Imaging; Radiation Oncology Physics; Semantics; Simulation; Spiral Cone-Beam Computed Tomography; Thorax; Thorax - diagnostic imaging; thorax CBCT; Tomography; X-rays; contextual loss; FFRN; artifacts correction; thorax CBCT
• ISSN: 1526-9914; 1526-9914
• DOI: 10.1002/acm2.13968

Background and objective
Cone‐beam computed tomography (CBCT) has become a more and more active cutting‐edge topic in the international computed tomography (CT) research due to its advantages of fast scanning speed, high ray utilization rate and high precision. However, scatter artifacts affect the imaging performance of CBCT, which hinders its application seriously. Therefore, our study aimed to propose a novel algorithm for scatter artifacts suppression in thorax CBCT based on a feature fusion residual network (FFRN), where the contextual loss was introduced to adapt the unpaired datasets better.
Methods
In the method we proposed, a FFRN with contextual loss was used to reduce CBCT artifacts in the region of chest. Unlike L1 or L2 loss, the contextual loss function makes input images which are not aligned strictly in space available, so we performed it on our unpaired datasets. The algorithm aims to reduce artifacts via studying the mapping between CBCT and CT images, where the CBCT images were set as the beginning while planning CT images as the end.
Results
The proposed method effectively removes artifacts in thorax CBCT, including shadow artifacts and cup artifacts which can be collectively referred to as uneven grayscale artifacts, in the CBCT image, and perform well in preserving details and maintaining the original shape. In addition, the average PSNR number of our proposed method achieved 27.7, which was higher than the methods this paper referred which indicated the significance of our method.
Conclusions
What is revealed by the results is that our method provides a highly effective, rapid, and robust solution for the removal of scatter artifacts in thorax CBCT images. Moreover, as is shown in Table 1, our method has demonstrated better artifact reduction capability than other methods.