Evaluation of a deep learning tool for detecting large vessel occlusion and intracranial hemorrhage on noncontrast computed tomography scans
Recommended Citation
Urra X, Rai A, Hernandez M, et al. Evaluation of a Deep Learning Tool for Detecting Large Vessel Occlusion and Intracranial Hemorrhage on Noncontrast Computed Tomography Scans. Stroke: Vascular and Interventional Neurology. 2025;5(6):e001872. doi:10.1161/SVIN.125.001872
Abstract
BACKGROUND: The purpose of this study is to assess the accuracy of automated artificial intelligence (AI) deep‐learning‐based modules in predicting suspected intracranial hemorrhage (ICH) or anterior circulation large vessel occlusion (LVO) on noncontrast computed tomography (NCCT) studies.
METHODS: We conducted a multicenter international retrospective cohort study, involving 6 stroke centers from the United States and Europe. We included patients in whom an acute stroke was suspected on admission and who underwent an NCCT and a CT angiography when ICH was not observed. Two neuroradiologists and a third one in case of discrepancies retrospectively evaluated all images and established the presence of ICH on NCCT and LVO on computed tomography angiography (ground truth). All NCCT scans were analyzed using 2 automated AI deep‐learning modules (Methinks, Barcelona, Spain) to assess the presence of ICH or LVO.
RESULTS: To assess the performance of the NCCT‐ICH module, a total of 200 patients were included in the study. The neuroradiologist's evaluation confirmed the presence of ICH in 97 cases (48.5%). To assess the performance of the NCCT‐LVO module, 382 patients were analyzed, with the neuroradiologist identifying a LVO in 141 cases (36.9%). The AI module for NCCT‐ICH detection demonstrated a sensitivity of 94.9% (95% CI]: 88.4%–98.3%) and specificity of 99.0% (95% CI: 94.7%–99.9%) with an area under the receiver operating characteristic curve of 0.974 (95% CI: 0.94–0.99). The LVO detection AI module on NCCT demonstrated a sensitivity of 81.6% (95% CI: 74.2–87.6), and specificity of 87.1% (95% CI: 82.2–91.1) with an area under the receiver operating characteristic of 0.915 (95% CI: 0.88–0.94).
CONCLUSIONS: The AI modules demonstrated high sensitivity and specificity in predicting ICH and LVO, suggesting their potential in offering support in clinical decisions in stroke networks immediately after NCCT is performed.
Document Type
Article