Automatic vessel segmentation in X-ray angiogram using spatio-temporal fully-convolutional neural network. (July 2021)

• Record Type: Journal Article
• Authors: Wan, Tao; Chen, Jianhui; Zhang, Zhonghua; Li, Deyu; Qin, Zengchang
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Vessel segmentation -- Spatio-temporal fully-convolutional neural network -- Conditional random field -- Coronary angiogram; 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.2021.102646
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A new deep learning based method for dynamic vessel segmentation is presented. A spatio-temporal FCN model provides accurate vascular structures with tiny vessel branches. The method can serve as a generalized framework to handle different image modalities. Abstract: Vessel segmentation from X-ray coronary angiogram (CAG) is essential in computer-aided diagnosis of cardiovascular diseases. Automatic segmentation is a challenging task due to the complex vascular structures and poor quality of CAG images. A new deep learning method is presented to automatically extract coronary arteries from dynamic CAG sequences. A spatio-temporal fully-convolutional neural network (ST-FCN) is designed to provide an effective way for segmenting entire vessel trees from motion sequences. An improved post-processing method subsequently refines the segmentation results by making a good use of spatial connectivity and temporal coherence between the moving CAG images. The ST-FCN model outperformed the state-of-the-art segmentation methods with dice similarity coefficient (DSC) of 0.90, accuracy (AC) of 0.92, and sensitivity (SN) of 0.89. Moreover, the ST-FCN achieved superior results in the stenosis detection task with AC of 0.95, SN of 0.92, specificity of 0.95, F1-score of 0.90 among all the reference approaches. The experimental results demonstrated that the integration of spatial and temporal information into a deep learning framework could enhance the vessel segmentation and might be useful for early detection of cardiovascular diseases.