Topology-Preserving Automatic Labeling of Coronary Arteries via Anatomy-Aware Connection Classifier

Zhixing Zhang, Ziwei Zhao, Dong Wang, Shishuang Zhao, Yuhang Liu, Jia Liu, Liwei Wang

Abstract

Automatic labeling of coronary arteries is an essential task in the practical diagnosis process of cardiovascular diseases. For experienced radiologists, the anatomically predetermined connections are important for labeling the artery segments accurately, while this prior knowledge is barely explored in previous studies. In this paper, we present a new framework called TopoLab which incorporates the anatomical connections into the network design explicitly. Specifically, the strategies of intra-segment feature aggregation and inter-segment feature interaction are introduced for hierarchical segment feature extraction. Moreover, we propose the anatomy-aware connection classifier to enable classification for each connected segment pair, which effectively exploits the prior topology among the arteries with different categories. To validate the effectiveness of our method, we contribute high-quality annotations of artery labeling to the public orCaScore dataset. The experimental results on both the orCaScore dataset and an in-house dataset show that our TopoLab has achieved state-of-the-art performance.

1. Introduction

In cardiovascular medicine, radiologists utilize Coronary Computerized Tomography Angiography (CCTA) to identify blockages. A critical but tedious step in this process is labeling each artery segment, such as the Left Anterior Descending (LAD) artery or the Right Coronary Artery (RCA). While deep learning models have automated much of this, they frequently produce "anatomically impossible" mistakes—such as predicting a branch of the left-side system as part of the right-side system.

These errors occur because standard models treat segments as independent objects, ignoring the category topological tree: the fixed, tree-like roadmap of how human heart arteries are physically connected. TopoLab aims to bridge this gap by forcing the AI to respect these anatomical rules through a topology-preserving design.

Vessel labeling results (a), category topological trees (b), and examples of predictions violating anatomical topologies (c).

2. Methodology

The TopoLab framework processes vessel segments through a three-stage pipeline designed to extract hierarchical features and enforce anatomical consistency:

Hierarchical Feature Extraction

  • Intra-segment Aggregation: Instead of using sequential models like LSTMs, which struggle with varying vessel lengths, TopoLab employs a Transformer encoder with a "segment query" to aggregate features along a single vessel segment.
  • Inter-segment Interaction: A Graph Convolutional Network (GCN) allows neighboring segments to communicate, capturing the spatial relationships and branching structure of the arterial tree.

Anatomy-aware Connection Classifier (AC-Classifier)

This core innovation shifts the task from labeling individual segments to classifying connected pairs of segments. The model compares segment pairs against "ground truth templates" derived from human anatomy (e.g., the Left Main artery must connect to the LAD). By classifying connections rather than isolated segments, the model is topology-preserving by design.

The TopoLab architecture, illustrating the flow from CCTA image encoding to hierarchical feature interaction and final AC-classification.

3. Experimental Results

The study utilized two primary datasets: the public orCaScore (72 CTA scans) and a large in-house dataset (1,200 CTA scans). The authors contributed new, high-quality anatomical annotations for 14 vessel classes to the orCaScore dataset to facilitate better benchmarking.

TopoLab outperformed existing state-of-the-art methods (like CPR-GCN and TreeLab-Net) across several metrics:

  • Accuracy: On orCaScore, TopoLab achieved an F1 score of 87.23%, a significant improvement over previous benchmarks (~83%).
  • Topology Preservation: The rate of topology violations (viola) dropped to 1.52%, compared to 3.75%–10.87% for competing models.
  • Case-Level Consistency: On the in-house dataset, the percentage of cases containing at least one topological error (viola-c) was slashed to 9.40%, whereas other methods ranged between 32% and 65%.

4. Conclusion

TopoLab demonstrates that medical AI performance can be significantly enhanced by explicitly constraining models with known biological structures. By framing labeling as a connection-classification task, the model naturally filters out anatomically impossible predictions. Future work may extend this "clinical priority" approach to other complex tubular structures, such as the airway or peripheral vascular systems.

Summary updated

Abstract

Automatic labeling of coronary arteries is an essential task in the practical diagnosis process of cardiovascular diseases. For experienced radiologists, the anatomically predetermined connections are important for labeling the artery segments accurately, while this prior knowledge is barely explored in previous studies. In this paper, we present a new framework called TopoLab which incorporates the anatomical connections into the network design explicitly. Specifically, the strategies of intra-segment feature aggregation and inter-segment feature interaction are introduced for hierarchical segment feature extraction. Moreover, we propose the anatomy-aware connection classifier to enable classification for each connected segment pair, which effectively exploits the prior topology among the arteries with different categories. To validate the effectiveness of our method, we contribute high-quality annotations of artery labeling to the public orCaScore dataset. The experimental results on both the orCaScore dataset and an in-house dataset show that our TopoLab has achieved state-of-the-art performance.
BibTeX
@inproceedings{zhang2023topology,
  author = {Zhixing Zhang and Ziwei Zhao and Dong Wang and Shishuang Zhao and Yuhang Liu and Jia Liu and Liwei Wang},
  title = {{Topology-Preserving Automatic Labeling of Coronary Arteries via Anatomy-Aware Connection Classifier}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  doi = {10.48550/arxiv.2307.11959},
}
Figure