Correlation study between quantitative characteristics of CT lung opacification based on machine learning and clinical subtypes and severity of lung injury of COVID-19 / 中华放射学杂志

ABSTRACT

Objective:To investigate the value of chest CT quantitative index in clinical classification and lung injury severity evaluation of COVID-19.Methods:The current study retrospectively analyzed the clinical and CT data of 438 patients with COVID-19 between January 2020 and March 2020 in Tongji Hospital, Tongji Medical College, Huazhong University of Science & Technology. The clinical types included common type ( n=146), severe type ( n=247) and critical type ( n=45). The chest CT indexes of all patients were quantitatively analyzed by artificial intelligence (AI) deep learning, including whole lung volume, CT lung opacification, ground glass opacification volume (GGO volume; CT value<-300 HU), solid opacification volume (SO volume; CT value ≥-300 HU) and the ratio of volume to the whole lung volume, the ratio of SO volume to GGO volume (SO volume/GGO volume). Kruskal-Wallis test was used to conduct statistical analysis of the differences in quantitative parameters among clinical types, and multiple ordered logistic regression was used to analyze the correlation between quantitative parameters and clinical types. Results:Among the 438 patients diagnosed with COVID-19, severe and critical patients were older ( P<0.05), and most of the critical patients were male ( P<0.05). The main clinical manifestations of all clinical types were fever, followed by cough, fatigue, chest tightness, dyspnea, gastrointestinal symptoms and so on. GGO volume was the main CT manifestation of all the three clinical subtypes. The whole-lung opacification volume, GGO volume, SO volume and their proportions in whole-lung volume significantly increased from common, severe to critical types (all P<0.05). SO volume/GGO volume increased with the severity of clinical type [common type 0.12 (0.03, 0.34), severe type 0.29 (0.11, 0.59), critical type 0.61 (0.39, 0.97)]. Multiple ordered logistic regression analysis showed that whole-lung opacification volume (OR=1.009), SO volume/GGO volume (OR=1.866), GGO volume (OR=1.008) and SO volume (OR=1.016) had a significant positive effect on the severity of clinical typing ( P<0.01). Conclusion:Quantitative indicators of chest CT based on deep learning algorithm (SO volume, GGO volume, SO volume/GGO volume) are closely related to the clinical severity of COVID-19.