Precise vascular pose estimation is critical for neurological workflows, offering enhanced intraoperative guidance and providing vascular landmarks that propagate spatial consistency to surrounding brain regions. However, current methods are hindered by feature occlusion from limited view, intricate anatomical structures and the scarcity of annotated surgical data. To address these challenges, we propose a graph-based vascular pose estimation method that leverages the vessel topology visible through cranial windows to achieve precise estimates and maintain spatial consistency. Our approach transforms vessels into topological graphs using centerline-guided node sampling and introduces structure-aware EdgeDrop layers (SA-EdgeDrops) that dynamically adjust edge connections based on vessel node types, which effectively preserves critical topological features and prevents over-smoothing in deep Graph Neural Networks. Meanwhile, we developed a synthetic dataset derived from patients’ medical images, incorporating vessels with structural complexity variations. This enables benchmarking and validation of pose estimation methods in anatomically challenging scenarios. Experimental results demonstrate that our method achieves better performance across vessels of varying geometries and sizes, yielding a 9.8% improvement in the ADD metric over FFB6D. Ablation studies validate the effectiveness of centerline-guided graph and SA-EdgeDrops. The dataset will be made available at https://github.com/WHHHHY/Structure-aware-GNNs-for-vascular-registration-in-AR-assisted-neurosurgery .
Hong et al. (Mon,) studied this question.