Automated Intracranial Artery Labeling Using a Graph Neural Network and Hierarchical Refinement

Li Chen, Thomas S. Hatsukami, Jenq-Neng Hwang, Chun Yuan

Introduction

Mapping the complex network of intracranial arteries (ICA) is a cornerstone of modern neurology, essential for diagnosing conditions ranging from acute strokes and aneurysms to long-term cognitive decline. At the center of this network lies the Circle of Willis (CoW), a critical vascular ring that provides redundant blood flow to the brain. Despite its importance, automated labeling of these vessels remains a significant challenge due to high anatomical variability—only about 52% of the population possesses a "complete" CoW—and pathological changes like stenosis that can obscure vessels or create confusing collateral pathways.

Overview of intracranial artery labeling: (a) raw MRA scan, (b) 3D reconstruction with automated labels, and (c) anatomical schematic of the Circle of Willis.

Traditional automated methods often rely on rigid templates that struggle with non-standard anatomy. Furthermore, the field has historically been limited by small datasets, hindering the development of robust, generalizable machine-learning models. This paper introduces a high-speed, high-accuracy system that combines Graph Neural Networks (GNN) with a novel Hierarchical Refinement process to overcome these hurdles.

Methodology

The proposed system models the brain's arterial tree as a mathematical graph, where bifurcations and endpoints serve as nodes and artery segments represent edges.

Graph Neural Network (GNN) Architecture

The core of the system is a message-passing GNN. Unlike models that analyze junctions in isolation, this GNN performs 10 rounds of message passing, allowing each node to incorporate context from its neighbors. To capture the essential 3D spatial relationships of the vasculature, the authors utilized a multi-label binary encoding representing 26 major directions in 3D space. This allows the model to learn that specific arteries typically follow predictable directional paths relative to the brain's axes.

Hierarchical Refinement (HR)

To ensure anatomical consistency that a pure deep learning approach might miss, the researchers implemented a three-level Hierarchical Refinement layer. This process mimics human anatomical reasoning by enforcing structural logic over local predictions:

  1. Anchor Identification: The system first identifies "confident" nodes where GNN predictions for both nodes and connected edges are in agreement.
  2. Sub-tree Construction: Using these confident anchors, the model builds the major arterial sub-trees (left anterior, right anterior, and posterior).
  3. Variable Branch Integration: Finally, optional or highly variable segments, such as the communicating arteries, are added based on the established structural framework.
The Hierarchical Refinement process, showing the step-by-step construction of the arterial tree from confident nodes to full sub-trees and variable branches.

Experimental Setup and Datasets

The study utilized a large-scale dataset of 729 Magnetic Resonance Angiography (MRA) scans, significantly larger than the typical ~50-scan datasets used in previous research. This cohort included 508 scans for training, 116 for validation, and 105 for testing, encompassing both healthy individuals and patients with vascular pathologies like hypertension and stroke. To evaluate generalizability, the model was also tested on the public UNC dataset without additional training.

Key Results

The system demonstrated exceptional performance across several metrics:

  • High Accuracy: The model achieved a 97.5% node labeling accuracy on the primary test set.
  • Completeness: In 63.8% of test cases, the system correctly labeled every single node in the Circle of Willis, a massive improvement over the ~5% success rate of traditional template-matching methods.
  • Efficiency: The GNN-based approach processes a scan in just 0.09 seconds, compared to over 8 minutes for some optimization-based predecessors, making it suitable for real-time clinical applications.
  • Robustness and Generalization: The model maintained high accuracy (99%) on the external UNC dataset and successfully handled challenging cases with missing segments or anatomical noise.

Discussion and Conclusion

By integrating the local feature-learning capabilities of GNNs with the global structural logic of Hierarchical Refinement, this research provides a robust solution for automated intracranial artery labeling. The "anatomy as logic" approach ensures that the model adheres to biological constraints while remaining flexible enough to handle natural variation.

A current limitation is the system's dependence on pre-processed artery "tracings" (centerlines). Errors in the initial segmentation or tracing can propagate through the graph construction. Future research will focus on developing an end-to-end pipeline that integrates vessel segmentation and labeling into a single unified process, further streamlining the clinical workflow.

Summary updated

BibTeX
@inproceedings{chen2020automated,
  author = {Li Chen and Thomas S. Hatsukami and Jenq-Neng Hwang and Chun Yuan},
  title = {{Automated Intracranial Artery Labeling Using a Graph Neural Network and Hierarchical Refinement}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  doi = {10.1007/978-3-030-59725-2\_8},
}
Figure