Automated CTA-Derived Collateral Grading and Morphologic Metrics for Enhanced Prediction of Post-Stroke Outcomes.
Authors
Affiliations (2)
Affiliations (2)
- From the University of California (A.D., K.L.), Riverside, CA; Division of Vascular Neurology and Neurocritical Care (J.W., K.M.B., L.A., S.Y., Z.B., D.M., G.K.H., P.T.-F.), Inova Neuroscience and Spine Institute, Inova Fairfax Medical Campus (IFMC), Falls Church, VA; Department of Medical Education (J.W., K.M.B., P.O., G.K.H., P.T.-F.), University of Virginia, IFMC Campus, Falls Chruch, VA; and Division of Neuroradiology (K.M.B., P.O.), IFMC, Falls Church, VA.
- From the University of California (A.D., K.L.), Riverside, CA; Division of Vascular Neurology and Neurocritical Care (J.W., K.M.B., L.A., S.Y., Z.B., D.M., G.K.H., P.T.-F.), Inova Neuroscience and Spine Institute, Inova Fairfax Medical Campus (IFMC), Falls Church, VA; Department of Medical Education (J.W., K.M.B., P.O., G.K.H., P.T.-F.), University of Virginia, IFMC Campus, Falls Chruch, VA; and Division of Neuroradiology (K.M.B., P.O.), IFMC, Falls Church, VA. [email protected].
Abstract
Collateral circulation is a key determinant of treatment response and outcomes in acute ischemic stroke (AIS), yet its assessment in clinical practice remains limited and subjective. While CT perfusion (CTP) offers insight into tissue viability, its restricted availability and susceptibility to artifacts reduce its practical utility, particularly in smaller centers. As an accessible alternative, we developed and validated an automated quantitative collateral index (qCI) derived from CTA using a deep learning U-Net segmentation framework, and evaluated the ability of CTA-based features to predict post-stroke recovery and functional outcomes. We retrospectively analyzed prospectively collected data from 230 AIS patients who underwent endovascular thrombectomy (EVT) between 2019-2023. CTA scans were segmented using a validated neural network-based vascular extraction model to generate 3D vessel networks and compute morphology metrics (vessel length, branching, fractal dimension, tortuosity). A fully automated qCI was derived through hemispheric comparison of vascular features following spatial registration. Agreement of qCI with clinician grading was quantified. Gradient-boosted decision tree models were trained to predict early neurological deterioration (END), early neurological improvement (ENI), and 90-day modified Rankin Scale (mRS) using (i) CTP-only (core, penumbra, mismatch), (ii) CTA-only (qCI + morphology), and (iii) combined CTA+CTP features. Automated qCI (graded 0-3) showed strong concordance with expert scoring (accuracy 0.863; Pearson R = 0.880; Cohen's Îș = 0.786). Dichotomized collateral status achieved 0.938 accuracy (AUROC = 0.945). For 90-day mRS prediction, the CTA-only model outperformed the CTP-only model (AUROC 0.730 vs 0.645) with better calibration (Brier score 0.178 vs 0.295). The combined CTA+CTP model performed best overall (AUROC 0.781), with similar improvements observed for END. CTA-derived features led to significant reclassification gains when added to perfusion-based models. Automated CTA-derived qCI and cerebrovascular morphology provide rapid, objective, and reproducible collateral assessment with high agreement to expert grading. These features outperform perfusion metrics in several predictive tasks and further enhance prognostic accuracy when combined with CTP. Because CTA is widely available, qCI offers a scalable, clinically practical tool for improving stroke outcome prediction, particularly in settings where CTP is unavailable.