Inter-rater reliability for assessing intracranial collaterals in patients with acute ischemic stroke: comparing 29 raters and an artificial intelligence-based software

• Published in: Neuroradiology 2022-12, Vol.64 (12), p.2277-2284
• Main Authors: Wolff, Lennard, Su, Jiahang, Van Loon, Derek, van Es, Adriaan, van Doormaal, Pieter Jan, Majoie, Charles, van Zwam, Wim, Dippel, Diederik, van der Lugt, Aad, van Walsum, Theo
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial intelligence; Automation; Blood circulation; Blood flow; Cardiovascular system; Correlation coefficient; Correlation coefficients; Diagnostic Neuroradiology; Evaluation; Imaging; Ischemia; Medicine; Medicine & Public Health; Neurology; Neuroradiology; Neurosciences; Neurosurgery; Occlusion; Patients; Radiology; Reliability analysis; Reproducibility; Stroke; Training
• ISSN: 0028-3940; 1432-1920
• DOI: 10.1007/s00234-022-02984-z

Purpose Outcome of endovascular treatment in acute ischemic stroke patients is depending on the collateral circulation maintaining blood flow to the ischemic territory. We evaluated the inter-rater reliability and accuracy of raters and an automated algorithm for assessing the collateral score (CS, range: 0–3) in acute ischemic stroke patients. Methods Baseline CTA scans with an intracranial anterior occlusion from the MR CLEAN study ( n =500) were used. For each core lab CS, ten CTA scans with sufficient quality were randomly selected. After a training session in collateral scoring, all selected CTA scans were individually evaluated for a visual CS by three groups: 7 radiologists, 13 junior and 9 senior radiology residents. Two additional radiologists scored CS to be used as reference, with a third providing a CS to produce a 2 out of 3 consensus CS in case of disagreement. An automated algorithm was also used to compute CS. Inter-rater agreement was reported with intraclass correlation coefficient (ICC). Accuracy of visual and automated CS were calculated. Results 39 CTA scans were assessed (1 corrupt CTA-scan excluded). All groups showed a moderate ICC (0.689-0.780) in comparison to the reference standard. Overall human accuracy was 65± 7% and increased to 88± 5% for dichotomized CS (0–1, 2–3). Automated CS accuracy was 62%, and 90% for dichotomized CS. No significant difference in accuracy was found between groups with different levels of expertise. Conclusion After training, inter-rater reliability in collateral scoring was not influenced by experience. Automated CS performs similar to residents and radiologists in determining a collateral score.