Abstract WP302: Validation of an AI-Powered Module for Rapid, Accurate 3D Cerebrovascular Segmentation in CTA: Impact on Workflow and Clinical Decision-Making

Abstract

Introduction: Cerebrovascular pathologies require rapid and accurate imaging interpretation to guide time-sensitive decisions in acute neurovascular care. Advanced 3D segmentation of CTA images offers the potential to accelerate diagnosis and improve communication between clinicians. RapidAI's Lumina 3D TM module automatically processes contrast-enhanced CT-Angiography (CTA) scans from the aortic arch to the cerebrum, delivering bone- and vein-subtracted vascular visualizations, including 3D volume renders (VR), rotational maximum intensity projection (MIPs), and 2D curved planar reformats (CPRs), to support stroke triage and treatment planning. This study aimed to determine the accuracy of Lumina 3D TM segmentations by comparing its segmentations against those of expert neuroradiologists. Methods: Fifty CTA scans with diverse cerebrovascular anomalies, including aneurysms, stenoses, atresia, dolichoectasia, and tortuous vessels, were manually segmented by 3 expert annotators and reviewed by a U.S board-certified neuroradiologist to establish gold standard 3D segmentations. Lumina 3D TM ’s automated outputs were compared using Dice coefficients and Hausdorff distances across extracranial and intracranial vascular regions. Ground truth reproducibility was evaluated by comparing multiple annotators’ segmentations of the cases. CPR centerline accuracy was also assessed. Median image generation time was recorded. Results: For extracranial vessels, Lumina 3D TM achieved a mean Dice Coefficient of 0.89 (95% CI: 0.87-0.90) and Hausdorff Distance of 0.44 mm (95% CI: 0.40-0.49) when compared to expert-reviewed manual segmentations. Intracranially, the module demonstrated superior accuracy to expert segmentations (Dice: 0.97 [95% CI: 0.97-0.98]; Hausdorff: 0.44 mm [95% CI: 0.37-0.52]. CPR centerline alignment showed a mean Hausdorff Distance of 0.31 mm (95% CI: 0.30-0.32). Ground truth reproducibility showed only 1% variance. The median time to generate Lumina 3D TM images was under 2 minutes. Conclusion: RapidAI’s Lumina 3D TM software produces clinically accurate, reproducible vascular segmentations with rapid turnaround, supporting efficient CTA interpretation in high-stakes stroke workflows. Its high fidelity to expert annotations and speed make it a valuable tool for clinical decision-making and streamlining neurovascular care.

Ähnliche Arbeiten