Automated segmentation of ruptured and non-ruptured abdominal aortic aneurysms based on active learning

Read full text Summarise with VIRAT
Authors: Guo, Jia, Fabien Lareyre, Sébastien Goffart, Martin Teraa, Regent Lee, Juliette Raffort, Hervé Delingette
Cited by: 0
References: 47

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Comprehensive and accurate anatomical characterization of abdominal aortic aneurysms (AAA) is essential for diagnosis, risk stratification and endovascular aneurysm repair planning. However, the development of robust segmentation models is limited by scarce ground truth annotations and heterogeneous annotation protocols. This study aimed to develop an active learning-assisted framework for comprehensive multi-class AAA segmentation under heterogeneous annotation conditions. A total of 175 multi-center computed tomography angiography scans were collected for model development, including scans from patients with ruptured AAA, patients with non-ruptured AAA, and healthy subjects with varying degrees of annotation completeness. The framework integrates a Dice similarity coefficient (DSC)-based active-learning (AL) with cross-validation based segmentation initialization to enable efficient utilization of partially labeled and unlabeled data. The AL-assisted correction process resulted in an estimated reduction of approximately 65% in annotation effort compared with conventional manual or semi-automatic annotation. Experimental results demonstrated strong segmentation performance across major vascular structures and proximal aortic branches, achieving a DSC of 0.93±0.09 for the aortic lumen, 0.75±0.18 for calcification, and 0.89±0.09 for intraluminal thrombus. For proximal branches, DSC values exceeded 0.88 for the celiac trunk, superior mesenteric artery, and the renal arteries. The proposed framework maintained robust performance across multi-center datasets and different clinical subgroups without evidence of systematic degradation in ruptured cases. Overall, this study demonstrates the feasibility and practical value of label-efficient learning for large-scale AAA segmentation and provides a scalable solution for multi-center vascular imaging applications.

Fields of Study & Subjects

Medical Image Segmentation Techniques Aortic aneurysm repair treatments Medical Imaging and Analysis

Publication details

Journal / source
HAL
Publication date
17 March 2026
Language
EN
Document type
Preprint