Abstract
It is essential to quantitatively analyze coronary arterial angiography (CAG) images for coronary interventional procedures as they are the gold reference standard for assessing stenosis. However, mainstream methods cannot extract different vascular branches precisely due to inadequate annotated data by experts and difficulty in uniformly defining vascular branches. To address the issue, this paper presents a novel framework to analyze multi-branching vessels with simple annotation, which localizes the endpoints of the vascular branches first, and then extracts corresponding vessel segments based on the endpoints. Considering it is challenging for current image-based methods to distinguish sparse vessel keypoints, we propose a cross-modal Language-Enhanced Endpoint Detection (LEED) network to achieve better performance by providing linguistic knowledge obtained from clinicians. A medical term corpus is established to convert conversation contents to standard texts. Furthermore, we devise a Twin-Stream Vessel Segmentation (TS-Seg) algorithm to segment various vascular branches based on gray-level features and detected endpoints. Besides, clinicians are allowed to manually add new endpoints and apply the TS-Seg algorithm alone to analyze any variant vascular branch that has not been defined before. Extensive experiments on two in-house and one public datasets consistently demonstrate the effectiveness and superiority of the proposed framework. Moreover, comprehensive vessel analysis reports are provided based on clinical prior knowledge to assist in preoperative surgery planning and postoperative treatment assessment, showing great potential to improve the efficiency and accuracy of clinical workflow.
Funding Information
  • National Key Research and Development Program of China (2024YFF1206902)
  • National Natural Science Foundation of China (62303463)
  • Beijing Natural Science Foundation (L232137)
  • Youth Innovation Promotion Association of CAS (2020140)