Proposing a New Standard for Collateral Status Assessment in Acute Ischemic Stroke using Cerebrovascular Radiomics

Abstract Background and Aims One of the major determinants of individual functional outcome following a large vessel occlusion (LVO) stroke is the degree of collateral circulation, mostly evaluated on CT angiography (CTA). Many clinical trials have established its importance in selecting patients for endovascular thrombectomy. However, assessment of collateral circulation in acute LVO using CTA is rater-dependent and prone to diagnostic errors. We developed an automated cerebrovascular radiomics pipeline to establish objective collateral scoring. Methods We retrospectively analyzed admission single-phase CTAs from 343 patients included in the MR CLEAN trial. The data was split and stratified by the collateral-score label into training and validation (n=274) and internal testing (n=69) sets. Vessel information was derived from two regions of interest used for radiomics feature extraction: 1) whole cerebral arterial tree, and 2) Circle of Willis artery segments. Segmentation models were developed using the nnU-Net framework on annotated CTAs of the cerebral arterial tree (n = 40) and multiclass Circle of Willis segmentations from the TopCoW dataset (n = 125), respectively. A customized feature selection pipeline identified the most predictive features, which were used to train a random forest classifier to predict collateral status (sufficient: >50% vs. insufficient: <50% filling). We compared the arteries-based approach to established atlas-based middle cerebral artery (MCA) masks and further validated it on an external dataset of 140 acute LVO patients. Results The vessel segmentation models accurately annotated cerebral arteries (95HD: 4.4932, average Dice coefficient: 0.8347) and the circle of Willis segments (95HD: 2.2711, average DICE coefficient: 0.8095). After radiomics selection, the best predictive features were identified for the cerebral vasculature (n = 6), the MCA mask (n = 98), and the combination model (n = 32). The Vessel-tree-based radiomics model outperformed the MCA mask approach on both internal (AUC: 0.8809 vs. 0.8160) and external (AUC: 0.8339 vs. 0.6628) test sets. Incorporating radiomics features from Circle of Willis segments further boosted performance, achieving the highest external test set AUC (0.8677). Conclusion We present an accurate, fully automated, and generalizable cerebrovascular radiomics approach for assessing collateral status from admission computed tomography angiography, supporting time-critical decision-making in acute large vessel occlusion.