Generating CT images in delayed PET scans using a multi-resolution registration convolutional neural network. (September 2022)

• Record Type: Journal Article
• Authors: Zhai, Mingwei; Yang, Yong; Sun, Fangfang; Wang, Xinhui; Wang, Xiaozhuang; Ke, Changjie; Yu, Chenhao; Ye, Hongwei
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: Positron emission tomography -- Computed tomography -- Image registration -- Convolutional neural network -- Multi-resolution framework; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103853
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlight: We proposed a novel registration network to generate DL-based delayed CT images. Abstract: A Delayed scan refers to the reacquisition of CT and PET after a regular scan to improve the sensitivity and specificity of PET/CT examination. However, an additional CT scan will lead much more X-ray radiation to the patient. Therefore, developing a method to generate delayed CT (T2CT) images to avoid additional CT scans is particularly important in clinic. This paper aims to generate T2CT images from delayed PET (T2PET), regular PET (T1PET) and regular CT (T1CT) images using deep learning methods. However, it may encounter difficulties such as intrinsic differences between multi-modal images and large deformations caused by two scans. To address these issues, a multi-resolution registration convolutional neural network (MRR-CNN) is introduced to improve the accuracy of generating CT images. MRR-CNN employs three models to separately predict deformation vector field (DVF) in different resolution levels. In this method, the global deformation is evaluated firstly, and then local deformations are gradually fused to generate accurate T2CT images. We selected a recently published deep learning-based method (VoxelMorph) to compare the effectiveness of our method on 10 clinical patient data, using mean absolute error (MAE) and root mean square error (RMSE) as evaluation metrics. Compared with VoxelMorph, the proposed MRR-CNN achieves lower MAE (61.26 vs. 67.24) and lower RMSE (118.74 vs. 126.13). The experimental results indicate that our proposed method outperforms VoxelMorph in generating T2CT images.