CE-NC-VesselSegNet: Supervised by contrast-enhanced CT images but utilized to segment pulmonary vessels from non-contrast-enhanced CT images. (April 2023)

• Record Type: Journal Article
• Authors: Wang, Meihuan; Qi, Shouliang; Wu, Yanan; Sun, Yu; Chang, Runsheng; Pang, Haowen; Qian, Wei
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Deep learning -- Pulmonary vessel segmentation -- Image registration -- Non-contrast computed tomography -- Contrast-enhanced computed tomography; 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.2022.104565
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Pulmonary vessel masks are accurately annotated in contrast-enhanced CT images. Vessel masks are transferred to non-contrast-enhanced CT images via registration. A model is gained to segment pulmonary vessels from non-contrast-enhanced CT images. The model enables more accurate and continuous segmentation of pulmonary vessels. Accurate annotations are the key to training a successful segmentation model. Abstract: Background: The automatic segmentation of pulmonary vessels from CT images has important significance. However, accurately annotating pulmonary vessels directly in non-contrast CT (NCCT) images is complex and time-consuming. Methods: This study aims to draw annotations with contrast-enhanced CT (CECT) images and train a deep-learning model for segmenting pulmonary vessels from NCCT images. Two datasets with 63 CT scans were collected. Dataset D1 included 17 cases annotated in CECT images, 10 cases annotated in NCCT images, and 12 NCCT scans. Dataset D2 consisted of 12 CECT and 12 NCCT scans with annotations. First, annotations drawn in CECT images (Dataset D1) are transferred to NCCT images via spatial registration. Second, a CE-NC-VesselSegNet is proposed and trained using the transferred annotations to segment pulmonary vessels from NCCT images. Finally, the CE-NC-VesselSegNet is evaluated and compared with its counterparts. Results: After registration, the maximum and root mean square error between CECT and NCCT images decreases, while the structural similarity and peak signal-to-noise ratio increase. CE-NC-VesselSegNet can accurately segment pulmonary vessels from NCCT images with a Dice of 0.856. In the external validation using Dataset D2, the CE-NC-VesselSegNet achieves a Dice of 0.738, which is higher compared with that of NC-VesselSegNet trained by D2. Visual inspections have shown that CE-NC-VesselSegNet enables more accurate and continuous segmentation compared with its counterpart. Conclusions: Annotations of pulmonary vessels drawn in CECT images can be transferred to NCCT images via spatial registration. Using these transferred annotations of high quality, a CE-NC-VesselSegNet can be trained to segment pulmonary vessels from NCCT images.