Abstract
Accurate coronary artery segmentation is critical for computeraided diagnosis of coronary artery disease (CAD), yet it remains challenging due to the small size, complex morphology, and low contrast with surrounding tissues. To address these challenges, we propose a novel segmentation framework that leverages the power of vision foundation models (VFMs) through a parallel encoding architecture. Specifically, a vision transformer (ViT) encoder within the VFM captures global structural features, enhanced by the activation of the final two ViT blocks and the integration of an attention-guided enhancement (AGE) module, while a convolutional neural network (CNN) encoder extracts local details. These complementary features are adaptively fused using a cross-branch variational fusion (CVF) module, which models latent distributions and applies variational attention to assign modality-specific weights. Additionally, we introduce an evidential-learning uncertainty refinement (EUR) module, which quantifies uncertainty using evidence theory and refines uncertain regions by incorporating multi-scale feature aggregation and attention mechanisms, further enhancing segmentation accuracy. Extensive evaluations on one in-house and two public datasets demonstrate that the proposed framework significantly outperforms state-of-the-art methods, achieving superior performance in accurate coronary artery segmentation and showcasing strong generalization across multiple datasets. The code is available at https://github.com/d1c2x3/CAseg.
Introduction: The Challenge of Coronary Segmentation
Coronary artery disease (CAD) is a leading cause of mortality globally. Clinicians rely on Coronary Computed Tomography Angiography (CCTA) to identify blockages, but automated segmentation of the coronary arteries is notoriously difficult. The task is complicated by three primary factors: the minute scale of the vessels relative to the heart volume, the significant intensity overlap between arteries and surrounding tissues (low contrast), and the complex, branching geometry of the vascular tree.
Traditional deep learning models often struggle with these constraints. Standard Convolutional Neural Networks (CNNs) are effective at capturing local textures but often lack the global context necessary to maintain vessel continuity, leading to fragmented results. Conversely, Vision Transformers (ViTs) excel at capturing long-range dependencies but may miss fine-grained edge details.
Methodology: Parallel Encoding and Variational Fusion
To bridge the gap between global context and local detail, the proposed framework employs a three-pronged architectural approach that "unleashes" the potential of 3D Vision Foundation Models (VFMs).
1. Parallel Encoding (ViT + CNN)
- ViT Branch: Utilizes a pre-trained 3D foundation model (SAM-Med3D) to capture global structural features. An Attention-guided Enhancement (AGE) module is integrated to focus on vascular continuity by activating the final ViT blocks.
- CNN Branch: A 3D UNet-style encoder extracts high-resolution local shapes and textures, providing the fine-grained detail the ViT might overlook.
2. Cross-branch Variational Fusion (CVF)
Rather than simple concatenation, the CVF module treats feature fusion as a variational problem, modeling features as Gaussian distributions. It employs variational attention to adaptively weight global versus local information, making the model more robust to the inherent noise and variability in medical imaging.
3. Evidential-learning Uncertainty Refinement (EUR)
This module identifies regions of high uncertainty—such as blurry vessel boundaries—using Subjective Logic theory. By generating an uncertainty map, the model can trigger a targeted refinement process that pulls in multi-scale features to "fix" ambiguous or disconnected segments.
Experimental Results and Key Insights
The framework was rigorously evaluated on three distinct datasets: CCTA119 (an in-house dataset of 119 volumes), ASOCA (40 scans from a MICCAI challenge), and ICAS-100 (100 scans from the ImageCAS dataset).
- State-of-the-Art Accuracy: On the CCTA119 dataset, the model achieved a Dice Similarity Coefficient (DSC) of 87.31%, a substantial improvement over the 81.65% achieved by a standard 3D-UNet.
- Precision and Connectivity: The Average Symmetric Surface Distance (ASSD) was reduced to 0.71mm, indicating superior alignment with actual vessel boundaries.
- Robust Generalization: The model demonstrated strong cross-dataset performance, achieving an 85.26% DSC on the ASOCA dataset after being trained on CCTA119, proving its robustness across different clinical environments.
Conclusion
This research demonstrates that large-scale pre-trained foundation models can be effectively adapted for specialized medical tasks when paired with local-feature specialists and intelligent fusion mechanisms. By treating uncertainty as a guide for refinement rather than a failure mode, the proposed framework significantly improves the accuracy and connectivity of coronary artery segmentations. Future work will focus on optimizing computational efficiency and further ensuring anatomical topological continuity.
Summary updated