Deep-learning model detects hyperdense artery sign on non-contrast CT in suspected stroke patients.

Computed tomography angiography (CTA) is the gold standard for detecting large vessel occlusion, but its acquisition and reconstruction delay time-critical workflow. The hyperdense artery sign (HAS) on non-contrast CT (NCCT) offers an immediate, albeit subtle, marker. We developed a fully automated deep-learning model for HAS detection and evaluated its utility as an adjunctive pre-CTA alert to support earlier workflow readiness while confirmatory vascular imaging is pending. Furthermore, we assessed the radiological validity of the model’s detections to ensure they correspond to genuine thrombi rather than artifacts. We trained a 3-step deep-learning pipeline (midline correction, ischemic core segmentation, HAS segmentation) on 690 NCCT scans. Clinical validation was performed in two complementary cohorts: Part 1A, a multicenter CSC triage cohort (n = 159) representing a workflow-enriched high-acuity setting, and Part 1B, a single-center consecutive all-comer suspected-stroke cohort (n = 226) representing a broader real-world population. The primary metric was the Positive Predictive Value (PPV) to assess the reliability of the alert as a workflow-support role. Technical validation was performed using a crossover multi-reader study (n = 10 specialists and residents) to evaluate whether AI-detected regions were radiologically perceivable by human readers. In Part 1A, the model achieved a sensitivity of 76.2% (80/105), specificity of 87.0% (47/54), accuracy of 79.9% (127/159), and PPV of 92.0% (80/87), indicating high reliability of positive alerts in a CSC triage setting. In Part 1B, the model achieved a sensitivity of 74.3% (26/35), specificity of 82.7% (158/191), PPV of 44.1% (26/59), NPV of 94.6% (158/167), and accuracy of 81.4% (184/226), reflecting preserved discrimination in a lower-prevalence, broader real-world population. In the reader study, model assistance significantly improved HAS-detection performance, increasing JAFROC Figure of Merit from 0.71 to 0.77 (p  The proposed model enables rapid HAS detection on NCCT and demonstrated complementary performance across two validation settings: high reliability of positive alerts in a workflow-enriched CSC triage cohort and preserved sensitivity/specificity in a broader consecutive cohort. These findings support its role as an adjunctive pre-CTA alert for earlier workflow readiness in high-probability settings, not as a stand-alone rule-out tool. The observer study further supports the radiological validity of the AI-highlighted regions.