The segmentation of the carotid vessel wall using black-blood magnetic resonance images was a crucial step in the diagnosis of atherosclerosis. The objective was to accurately isolate the region between the artery lumen and outer wall. Although supervised learning methods achieved remarkable accuracy in vessel segmentation, their effectiveness remained limited due to their reliance on extensive labeled data and human intervention. Furthermore, when confronted with three-dimensional datasets featuring insufficient and discontinuous label data, these learning-based approaches could lose their efficacy. In this paper, we proposed a novel Joint 2D–3D Cross-Pseudo Supervision (JCPS) method for accurate carotid vessel wall segmentation.
In this study, a vascular center-of-gravity positioning module was developed to automatically estimate the region of blood vessels. To achieve accurate segmentation, we proposed a joint 2D–3D semi-supervised network to model the three-dimensional continuity of vascular structure. In addition, a novel loss function tailored for vessel segmentation was introduced, consisting of four components: supervision loss, cross-pseudo supervision loss, pseudo label supervision loss, and continuous supervision loss, all aimed at ensuring the accuracy and continuity of the vessel structure. In what followed, we also built up a user-friendly Graphical User Interface based on our JCPS method for end-users.
Our proposed JCPS method was evaluated using the Carotid Artery Vessel Wall Segmentation Challenge dataset to assess its performance. The experimental results clearly indicated that our approach surpassed the top 10 methods on the leaderboard, resulting in a significant enhancement in segmentation accuracy. Specifically, we achieved an average Dice similarity coefficient increase from 0.775 to 0.806 and an average quantitative score improvement from 0.837 to 0.850, demonstrating the effectiveness of our proposed JCPS method for carotid artery vessel wall segmentation.
The experimental results suggested that the JCPS method had a high level of generalization performance by producing pseudo labels that were comparable with software annotations for data-imbalanced segmentation tasks.