.png){kind=icon}
.png){kind=icon}
Return
Self-supervised vessel segmentation in X-ray angiograms using a boundary encoder-decoder structure
.png){kind=icon}
PRE
AI
DOI：10.1007/s00521-026-12064-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Using computers, automatically segmenting X-ray coronary angiograms (CAG) is essential for identifying and planning treatments for diseases affecting the heart’s blood vessels. Achieving precise coronary artery segmentation in X-rays is costly and time-consuming for cardiologists due to uneven noise, poor lighting, and distracting background details that complicate accurate segmentation of the X-ray CAG. To address these challenges, a novel framework is proposed to segment a CAG tree without the involvement of cardiologists. The framework operates in two phases: establishing a vessel tree ground truth and enhancing the generated mask. First, we will establish a ground truth for X-ray CAG through image denoising, contrast enhancement, and border removal. Next, Vessel-Enhanced Boundary U-Net (VEB U-Net) enhances the segmented vessel tree by combining boundary features with the main X-ray CAG. A Sobel edge detector has been incorporated into the segmentation model to improve boundary feature representation. A novel loss function utilizes boundary and shape information to produce more accurate results. The experiments used a public dataset of 2291 grayscale CAG X-ray images from 100 patients. Based on evaluations, the framework achieves the following: (i) It removes the requirement for manual ground truth in a supervised model, (ii) it reduces the cost and time needed for cardiologists to label X-ray CAGs, and (iii) it improves the segmented vessel tree with an accuracy of 96.9% based on a test set assessed by a single expert cardiologist. After validation from the cardiologist, the experiment indicated that the system successfully separated the vessels and could help detect heart diseases early.
Keywords:
Vessel segmentation
Convolution Neural Network
Boundaries information
Image processing techniques
Journal
IF:
4.5
Papers:
819
Citations:
3.2W