Abstract Number ‐ 51: Automated Versus Human Hyperdense Vessel Sign Detection Using Non‐Contrast Computed Tomography Scans

Abstract

Introduction Rapid detection of large vessel occlusion (LVO) is very crucial in triaging stroke patients potentially candidates for mechanical thrombectomy (MT). Hyperdense vessel sign (HDVS) is one of the earliest ischemic changes in non‐contrast CT scan (NCCT) indicating LVO stroke. Artificial intelligence emerged to detect HDVS with the advantages of faster acquisition, less variation, and a lower need for experience than the usual detection. We aimed to identify the diagnostic performance of automated software (e‐Stroke, Brainomix) in HDVS detection. Methods A prospectively collectedMT database from March 2020 to August 2021 was reviewed. Patients were included if they had intracranial internal carotid artery or middle cerebral artery M1 or M2 occlusion. Cases with HDVS were identified through the routine 2.5‐mm slice thickness NCCT scans after being correlated with patients’ clinical information and confirmed with CT angiography (CTA) scans. NCCT scans were classified according to slice thickness into two groups: 2.5‐mm scans and 0.625‐mm generated scans. All NCCT scans were read by e‐Stroke software, then deidentified and reviewed by two stroke neurologists who were blinded to any clinical, other imaging, or therapeutic information. They were required to record the presence/laterality of HDVS before and after observing other NCCT early ischemic changes like gaze deviation, loss of insular ribbon, caudate or lentiform hypodensity. ROC curve analysis was used to estimate sensitivity and specificity and the area under the curve (AUC) was compared using DeLong’s test. Inter‐rater agreement between the two readers’ final reads, e‐Stroke, and the standard read was measured using the Fleiss Kappa test. Results Among 304 patients included in the study, 37.7% had HDVS. Approximately 44% of the scans had 2.5‐mm slice thickness and 56% had 0.625‐mm slice thickness. The e‐Stroke software identified HDVS with a sensitivity of 63% and a specificity of 71% (Table 1). The mean AUC value of e‐Stroke HDVS detection (0.67[0.61‐0.74]) was similar to reader‐1 (0.68[0.62‐0.74];p = 0.87) and reader‐2 (0.63[0.57‐0.70];p = 0.56). HDVS detection improved by reader‐1(0.78[0.72‐0.83];p = 0.03) after observing other early ischemic changes on the same scans, but reader‐2 performance remained similar to e‐Stroke (0.69[0.63‐0.76];p = 0.71). AUC, sensitivity and specificity ofHDVS detection by e‐Stroke were significantly higher using 2.5‐mm compared to 0.625‐mm sliced NCCT scans (0.78[0.70‐0.86],sensitivity 70%,specificity 86%;p< 0.001) vs (0.58[0.50‐0.67],sensitivity 56%,specificity 61%;p = 0.06) respectively;p = 0.01. The readers also had higher AUC values with 2.5‐mm scans but not statistically significant, (0.74[0.66‐0.83] vs 0.64[0.56‐0.73];p = 0.18) for reader‐1 and (0.68[0.59‐0.77] vs 0.57[0.48‐0.66];p = 0.23) for reader‐2. The same after the final read, (0.85[0.78‐0.92] vs 0.75[0.67‐0.82];p = 0.08) for reader‐1 and (0.73[0.65‐0.82] vs 0.67[0.58‐0.76];p = 0.43) for reader‐2. Similarly, inter‐rater agreement was higher using 2.5‐mm sliced scans, k = 0.50(0.43‐0.75) compared to0.625‐mm scans,k = 0.27(0.21‐0.33). Conclusions Artificial intelligence (e‐Stroke software) has comparable sensitivity and specificity to human readers in HDVS detection. For e‐Stroke software, 2.5‐mm sliced CT scans are better to identifyHDVS compared to 0.625‐mm scans.