Abstract WP76: Automated Detection of Hyperdense MCA Sign and Automated Notification of Large Vessel Occlusion Using Artificial Intelligence

Abstract

Introduction: The hyperdense vessel sign (HDVS) on non-contrast computed tomography (NCCT) is one of the earliest findings in patients with acute ischemic stroke. Further, the HDVS has been associated with poor outcomes. Recognition of the HDVS, which indicates the presence and location of a red blood cell-rich intra-arterial thrombus, aids the diagnosis of large vessel occlusions (LVOs) and helps the triage to endovascular reperfusion. An efficient workflow to identify patients eligible for reperfusion therapy is essential to achieve the best patient outcome, and an artificial intelligence technology that can automatically detect HDVS and notify specialists may streamline this process. We present the first results using artificial intelligence to accurately identify HDVS for LVO detection. Methods: A convolutional neural network (Viz HDVS) was developed to detect hyperdense middle cerebral artery (MCA) sign of the M1 segment on NCCT. A single center retrospective analysis of NCCT in consecutive acute ischemic stroke patients from January 2014 to May 2015 with angiography-proven LVO was performed. Results of the Viz HDVS algorithm were compared to findings reported by expert neuroradiologists in the electronic medical record database, which were independently reassessed for congruence by a separate radiologist. Patients were excluded if a pre-angiography NCCT or the radiology report was unavailable for review, or if the presence of hyperdense MCA M1 sign was indeterminate. Results: A total of 223 patients were reviewed. The Viz HDVS algorithm identified hyperdense MCA M1 sign with a sensitivity of 70%, a specificity of 96%, and an AUC of 0.85 in 117 NCCT scans from acute ischemic stroke patients with LVO. The mean algorithm run time was 1 minute 30 seconds, with a maximum run time of 3 minutes. Projected mean time to specialist notification using Viz HDVS was 5 minutes. Conclusions: The novel Viz HDVS algorithm demonstrated remarkable feasibility and speed of applying artificial intelligence in the automated detection of HDVS and notification of specialists. This algorithm may aid in earlier identification and treatment of acute ischemic stroke patients with LVOs.