A Joint Geometric Topological Analysis Network (JGTA-Net) for Detecting and Segmenting Intracranial Aneurysms

The rupture of intracranial aneurysms leads to subarachnoid hemorrhage. Detecting intracranial aneurysms before rupture and stratifying their risk is critical in guiding preventive measures. Point-based aneurysm segmentation provides a plausible pathway for automatic aneurysm detection. However, challenges in existing segmentation methods motivate the proposed work.

We propose a dual-branch network model (JGTA-Net) for accurately detecting aneurysms. JGTA-Net employs a hierarchical geometric feature learning framework to extract local contextual geometric information from the point cloud representing intracranial vessels. Building on this, we integrated a topological analysis module that leverages persistent homology to capture complex structural details of 3D objects, filtering out short-lived noise to enhance the overall topological invariance of the aneurysms. Moreover, we refined the segmentation output by quantitatively computing multi-scale topological features and introducing a topological loss function to preserve the correct topological relationships better. Finally, we designed a feature fusion module that integrates information extracted from different modalities and receptive fields, enabling effective multi-source information fusion.

Experiments conducted on the IntrA dataset demonstrated the superiority of the proposed network model, yielding state-of-the-art segmentation results (e.g., Dice and IOU are approximately 0.95 and 0.90, respectively). Our IntrA results were confirmed by testing on two independent datasets: One with comparable lengths to the IntrA dataset and the other with longer and more complex vessels.

The proposed JGTA-Net model outperformed other recently published methods (>10% in DSC), showing our model's strong generalization capabilities.

The proposed work can be integrated into a large deep-learning-based system for assessing brain aneurysms in the clinical workflow.