Graph convolutional networks for automated intracranial artery labeling

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neuroscience and neurology

Background:

• Unruptured intracranial aneurysms (UIAs) pose a risk of subarachnoid hemorrhage, a critical stroke type.
• Accurate identification of intracranial arteries is crucial for assessing UIA-related risk factors.
• Automated methods are needed to improve the efficiency and accuracy of intracranial artery labeling.

Purpose of the Study:

• To enhance intracranial artery labeling using graph convolutional networks (GCNs).
• To investigate the effectiveness of atlas-based features in improving GCN performance for artery classification.
• To compare different GCN operators for optimal performance in this task.

Main Methods:

• Utilized 3D time-of-flight magnetic resonance angiography scans from 150 individuals.
• Employed GCNConv and GraphConv operators within GCN models for classifying 12 arterial bifurcations.
• Applied cross-validation and Wilcoxon signed-rank tests to evaluate atlas-based features and model performance.

Main Results:

• Atlas-based features significantly improved node classification accuracy ($$).
• The GraphConv operator demonstrated superior performance with a mean recall of 0.87 and precision of 0.90.
• Excellent model calibration was observed with an expected calibration error of 0.02.

Conclusions:

• Incorporating atlas-based features enhances intracranial artery labeling accuracy.
• The GraphConv operator is recommended over GCNConv due to its ability to integrate higher-order structural information.
• Improved artery labeling facilitates better identification of risk factors for unruptured intracranial aneurysms.