Detection and segmentation of hyperdense middle cerebral artery sign on non-contrast CT using artificial intelligence

Abstract

Background: The hyperdense artery sign (HAS) in patients with large vessel occlusion (LVO) is associated with outcomes after ischemic stroke. Considering the labor-intensive nature of manual segmentation of HAS, we developed and validated an automated HAS segmentation algorithm on non-contrast brain CT (NCCT) images using a multicenter dataset with independent annotations by two experts. Methods: For the training dataset, we included patients with ischemic stroke undergoing concurrent NCCT and CT angiography between May 2011 and December 2022 from six stroke centers. The model was externally validated using a dataset from one stroke center. For the clinical validation dataset, a consecutive series of patients admitted within 24 hours of symptom onset were included between December 2020 and April 2023 from six stroke centers. The model was trained using a 2D U-Net algorithm with manual segmentation by two experts. We constructed models trained on datasets annotated individually by each expert, and an ensemble model using shuffled annotations from both experts. The performance of the models was compared using area under the receiver operating characteristics curve (AUROC), sensitivity, and specificity. Results: A total of 673, 365, and 774 patients were included in the training, external validation, and clinical validation datasets, respectively, with mean (SD) ages of 68.8 (13.2), 67.6 (13.4), and 68.8 (13.6) years and male frequencies of 55.0%, 59.5%, and 57.6%. The ensemble model achieved higher AUROC and sensitivity compared to the models trained on annotations from a single expert in the external validation dataset. In the clinical validation dataset, the ensemble model exhibited an AUROC of 0.846 (95% CI, 0.819?0.871), sensitivity of 76.8% (65.1?86.1%), and specificity of 88.5% (85.9?90.8%). The predicted volume of the clot was significantly correlated with infarct volume on follow-up diffusion-weighted imaging (r=0.42; p<0.001). Conclusion: Our algorithm promptly and accurately identifies clot signs, facilitating the screening of potential patients who may require intervention.