ABSTRACT NUMBER: ESOC2026A2202 FULLY AUTOMATED DEEP LEARNING CEREBROVASCULAR MORPHOMETRY VALIDATION AND BILATERAL SYMMETRY ANALYSIS IN ACUTE ISCHEMIC STROKE

Abstract Background and aims Cerebrovascular morphometry may influence mechanical thrombectomy outcomes, but manual measurements are impractical clinically. Contralateral vessels could serve as patient-specific references. We developed a fully automated deep learning pipeline for cerebrovascular landmark detection, validated morphometric feature extraction against expert annotations, and established bilateral symmetry reference ranges in healthy subjects and stroke patients. Methods A 3D U-Net was trained to automatically detect six anatomical landmarks (bilateral extracranial ICA, terminal ICA, and MCA bifurcations) on CTA. Seventeen morphometric features were extracted across three arterial segments. Symmetry indices comparing left-right vessels were computed. Automated-expert agreement was assessed using ICC and Bland-Altman analysis (n=301). Symmetry distributions were compared between controls (n=402) and stroke patients with large vessel occlusion (n=472). Outcome associations were explored using cross-validated machine learning. Results Validation demonstrated excellent agreement for mean diameter (ICC=0.91), curvature energy (ICC=0.91), and tortuosity index (ICC=0.87), with minimal bias (<0.1mm). Controls showed bilateral symmetry for diameter (mean symmetry index 0.9%; 87% within ±20%). Stroke patients exhibited significant asymmetry for length (p=0.002), distal diameter (p<0.001), and curvature metrics (p<0.001), consistent with occlusion effects. For eTICI 2c-3 prediction, contralateral morphometry outperformed ipsilateral features (AUC 0.57 [95%CI 0.48-0.66] vs 0.54 [0.45-0.63]). Conclusions Fully automated landmark-based cerebrovascular morphometry achieves expert-level accuracy. Bilateral symmetry analysis reveals significant differences between controls and stroke patients. Contralateral vessels show potential as patient-specific references for outcome prediction. Conflict of interest Jesus David Gonzalez: Nothing to disclose. Pere Canals: Nothing to disclose. Magda Jablonska: Nothing to disclose. Giulio Fiore: Nothing to disclose. Alvaro Garcia-Tornel: nothing to dislcose. Marc Ribo reports a relationship with Anaconda Biomed S.L. that includes: board membership. Marc Ribo reports a relationship with Nora that includes: board membership. Marc Ribo reports a relationship with Medtronic Inc that includes: consulting or advisory. Marc Ribo reports a relationship with Cerenovus that includes: consulting or advisory. Marc Ribo reports a relationship with Methinks that includes: consulting or advisory. Marc Ribo reports a relationship with Vesalio that includes: consulting or advisory. Marc Ribo reports a relationship with Stryker that includes: consulting or advisory. Marc Ribo reports a relationship with Philips that includes: consulting or advisory. Marc Ribo reports a relationship with Rapid Pulse that includes: consulting or advisory. Marc Ribo reports a relationship with Sensome that includes: consulting or advisory. Marc Ribo reports a relationship with Apta Targets that includes: consulting or advisory. Figure 1 - belongs to Conclusions