ACE-ProtoNet: Adaptive covariance eigen-gate and uncertainty-aware prototype learning for coronary artery segmentation.

Abstract

Accurate segmentation of coronary arteries from Coronary CT Angiography (CCTA) is essential for quantitative stenosis evaluation, plaque characterization, and surgical planning. However, low vessel-to-background contrast, high anatomical variability, and the complex, tree-like vascular morphology pose significant challenges to automated segmentation. To address these challenges, we propose ACE-ProtoNet, a unified framework that couples an Adaptive Covariance Eigen-Gate (ACE-Gate) with an Uncertainty-aware Prototype Learning Head (UPL-Head) to achieve robust and accurate coronary artery segmentation. The architecture is built upon a parallel dual-encoder backbone which synergizes a partially frozen Vision Foundation Model (VFM) for global structural encoding with a trainable CNN for fine-grained local feature extraction. To reconcile the heterogeneity between these feature streams, the ACE-Gate explicitly models inter-channel dependencies through covariance analysis and eigenvalue decomposition, yielding statistically grounded, channel-wise gating for principled feature integration. Meanwhile, the UPL-Head leverages voxel-wise predictive uncertainty to modulate prototype-guided attention and dynamically update prototypes during training, thereby enhancing representation robustness in hard-to-classify regions and improving overall segmentation accuracy. Extensive experiments on two in-house and four public datasets demonstrate that ACE-ProtoNet consistently outperforms twelve state-of-the-art methods across multiple metrics, exhibiting superior cross-domain generalization as well as strong cross-modality and cross-anatomy transferability. The code is available at .