Abstract WP97: Automated Analysis For Detecting Large Vessel Occlusion On Non Contrast Ct In Acute Stroke

Abstract

Background and Purpose: Rapid detection of large vessel occlusion (LVO) is essential for rapid triage for thrombectomy to salvage at-risk tissue. Early detection of LVO on noncontrast CT head (NCCT) could facilitate triage before advanced imaging is obtained. RAPID hyperdense vessel sign (HVS), a new fully automated software- tool for HVS detection is compared to experienced radiologists, for early LVO detection on NCCT, and validated with CTA. Methods: Scans were selected from a database of 335 scans, of which 166 were selected by stratified random selection. Images were evaluated by the Rapid automated software, and six readers, two radiologists, and four neuroradiologists. The Rapid algorithm returns a score between zero and one, and a threshold of 0.5 was applied of HVS. Radiologists read each scan twice, 30 days apart. A scan rating confidence scale was used, from “very confident of absence,” to “very confident of HVS,” and a separate “technically inadequate” rating, and this was compared to the Rapid HVS score. “Very Confident,” or “Confident” rated scans were considered HVS positive. Results: Demographics: 54% were men. Ages ranged 23 - 94 with a median of 65. GE, Philips, Siemens, and Toshiba scanners are represented. Agreement with RAPID for HVS detection was 75% to 84%. Below a HVS score of 0.4 the proportion of LVO positive scans was 46% (indeterminate), between 0.4 and 0.6 the proportion was 62.5% (probable), and above 0.6 the proportion was 90% (likely). The scan rating confidence grade correlated with the Rapid HVS score. The software identified 53% of the LVOs, comparable to human readers (26-57%) A clinical practice simulation of RAPID followed by a reader identified 61.3% (p=0.02 vs readers alone). Among the 61 patients with no LVO on CTA, readers had a false positive rate of less than 7% as compared to 15% for the software. Median software processing time was 54 seconds. Conclusions: Expert readers, and the Rapid software, detect an HVS in about 50% of CT head images with LVO on CTA. When Rapid software is used in tandem with an experienced reader, HVS can be detected in up to two thirds of patients with LVO on CTA. This high percentage of early LVO detection would facilitate faster triage of patients for more advanced imaging and/or therapeutic interventions.