Semi 3D-TENet: Semi 3D network based on temporal information extraction for coronary artery segmentation from angiography video. (August 2021)

• Record Type: Journal Article
• Authors: Liang, Dongxue; Wang, Lu; Han, Dewei; Qiu, Jing; Yin, Xiaolei; Yang, Zhiyun; Xing, Junhui; Dong, Jianzeng; Ma, Zhaoyuan
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Temporal information -- Coronary angiography -- Video segmentation -- Semi 3D; 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.2021.102894
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Coronary artery interventional therapy is a clinically effective minimally invasive surgery for coronary artery disease. Extracting effective coronary vascular structures from coronary angiography videos is essential for the safe navigation of coronary interventional equipment and for the doctor to observe the location of the lesion. This paper proposes a new semi 3D architecture that uses the temporal information of video to segment coronary arteries from angiography video. We combine the 3D U-Net and 2D U-Net through a dimension conversion layer and a context extracting module. The input of the 3D encoder is a set of coronary video sequences. After the extracted three-dimensional features pass through the dimension conversion layer and the context information extraction module, the valuable features are input into the 2D decoder module. Finally, a clearer and more complete coronary is extracted to help the doctor to observe the vascular status better. We tested this method and the comparison methods on the coronary angiography video data set we made before. We can see from the experimental results that even in coronary angiography video sequences with poor quality, our method can achieve better results than the other methods. The accuracy of our results can reach 98.60%, which shows that in the vessel video segmentation task, the extraction of temporal information is helpful to extract a more complete vascular structure.