The human-AI scoring system: A new method for CT-based assessment of COVID-19 severity

Abstract

BACKGROUND: Chest computed tomography (CT) plays an important role in the diagnosis and assessment of coronavirus disease 2019 (COVID-19). OBJECTIVE: To evaluate the value of an artificial intelligence (AI) scoring system for radiologically assessing the severity of COVID-19. MATERIALS AND METHODS: Chest CT images of 81 patients (61 of normal type and 20 of severe type) with confirmed COVID-19 were used. The test data were anonymized. The scores achieved by four methods (junior radiologists; AI scoring system; human-AI segmentation system; human-AI scoring system) were compared with that by two experienced radiologists (reference score). The mean absolute errors (MAEs) between the four methods and experienced radiologists were calculated separately. The Wilcoxon test is used to predict the significance of the severity of COVID-19. Then use Spearman correlation analysis ROC analysis was used to evaluate the performance of different scores. RESULTS: The AI score had a relatively low MAE (1.67–2.21). Score of human-AI scoring system had the lowest MAE (1.67), a diagnostic value almost equal to reference score (r= 0.97), and a strongest correlation with clinical severity (r= 0.59, p< 0.001). The AUCs of reference score, score of junior radiologists, AI score, score of human-AI segmentation system, and score of human-AI scoring system were 0.874, 0.841, 0.852, 0.857 and 0.865, respectively. CONCLUSION: The human-AI scoring system can help radiologists to improve the accuracy of COVID-19 severity assessment.