Automated registration of 3D neurovascular territory atlas to 2D DSA for targeted quantitative angiography analysis in subarachnoid hemorrhage

Subarachnoid hemorrhage (SAH), a life-threatening condition caused by intracranial aneurysms, requires precise imaging for effective interventions. While digital subtraction angiography (DSA) remains the gold standard for SAH assessment, its limited region-specificity can compromise diagnostic accuracy. This study introduces a method to integrate a 3D neurovascular atlas with 2D DSA images for targeted, region-specific quantitative analysis, tailored for left and right internal carotid artery (ICA) and posterior injection scenarios. This study analyzed 245 angiograms from 184 SAH patients, covering left ICA, right ICA, and posterior injections, with both frontal and lateral views, yielding 1,470 total image sequences. A 3D vascular atlas was co-registered to DSA sequences using affine transformations refined with B-spline registration, guided by C-arm metadata and perfusion masks. Forty-eight regions of interest (ROIs) were defined for each injection type, representing 24 arterial territories per view. Time-density curves (TDCs) were generated by averaging pixel intensities within each ROI as a function of time, interpolated using Pchip for smoothness. API parameters, including bolus arrival time (BAT), time to peak (TTP), mean transit time (MTT), peak height (PH), and area under the curve (AUC), were extracted. Normalization to inlet data reduced injection-induced variability, while cross-correlation assessed temporal similarity between ROIs and the inlet. Dimensionality reduction via principal component analysis (PCA) retained five components per injection type, which were analyzed using multinomial logistic regression to predict clinical outcomes (no DCI, DCI, and death). Deformable registration enabled precise alignment of the 3D atlas with 2D DSA, facilitating region-specific API analysis. PCA identified key components in each injection type, with moderate predictive accuracy for clinical outcomes. This method enhances diagnostic precision and provides a foundation for future integration with advanced machine learning models to support real-time decision-making in SAH management.