Development and Validation of Automated Software for the Detection of Large Vessel Occlusion on Noncontrast CT

BackgroundReliable detection of large-vessel occlusion (LVO) via medical-image analysis has significant advantages in cases of acute ischemic stroke (AIS). In recent years, convolutional neural network (CNN)-based technologies for automated LVO detection have been developed. However, the pros and cons of CNN-based assistance in clinical practice remain poorly understood. The purpose of this study was to develop and validate a deep learning-based model to detect the hyperdense-artery sign (HAS) as a proxy for LVO and to investigate its impact on neurosurgeons diagnostic accuracy. MethodsWe conducted a multicenter, retrospective study of patients with LVO due to anterior-circulation AIS who underwent computed tomography angiography or magnetic resonance angiography on admission, as well as patients without LVO (patients with no indicative angiography features and patients with cerebral infarction without LVO), who were admitted from 2006 to 2022. A CNN algorithm for LVO detection was developed using data from four institutions (n=690), and model performance was validated using data from five institutions (n=129). For further investigation, five board-certified and five non-board-certified neurosurgeons performed two separate observer-performance studies with a 4-week interval, with and without the CNN for each image. ResultsThe HAS was detected in the correct location with a sensitivity and specificity of 0.79 and 0.87 by the CNN, 0.61 and 0.60 by board-certified neurosurgeons, and 0.61 and 0.66 by non-board-certified neurosurgeons, respectively. With the CNN, the mean area under the curve and figure of merit significantly increased for all readers (from 0.72 to 0.81, p<0.001, and from 0.71 to 0.77, p=0.005, respectively). ConclusionsOur deep learning-based automated LVO-detection model for non-contrast-enhanced computed tomographic images significantly improved neurosurgeons diagnostic performance. Further studies are needed to clarify the usefulness of the CNN in clinical practice.