Analysis of Prognosis of Coronavirus Disease 2019 Using Quantitative Measurement of Deep Learning

Abstract

Objective: To analyze the differences of chest computed tomography (CT) inflammatory lesions in patients with coronavirus disease 2019 (COVID-19), which were quantitatively measured based on deep learning, to warn the occurrence of severe cases and improve the understanding of the prognosis of COVID-19. Methods: The chest CT scans of 477 local patients with COVID-19 diagnosed for the first time at Inner Mongolia Autonomous Region People's Hospital were retrospectively analyzed. A total of 276 men and 201 women were divided into group A (not serious) and group B (serious) based on whether their diseases turned serious (severe/critical). Comparison was made between the two groups on the basic CT signs, such as lesion distribution, involved lobe side, number, differences in lesion volume, volume proportion, and density based on deep learning. Results: All 477 patients with COVID-19 had epidemiological history, and no statistical difference was noted in age and gender between the two groups. The volume and proportion of the lesions in the whole lung and each lobe of the lung in group B were higher than those in group A. The lesions in group A were mainly in the lower lobe of the right lung, accounting for 3.32% more than that in other lobes. The lower lobe of the left lung was the next, accounting for 2.08%. The volume of lesions in the upper lobe of the left lung was lower than that in other lobes, accounting for only 0.25%. No lesions were noted in the upper lobe of the right lung, middle lobe of the right lung, and upper lobe of the left lung in part of group A. In group B, the lesions were distributed in both lungs and in each lung lobe. The lower lobes of the right lung and left lung were predominant, accounting for 57.86% and 54.76%, respectively. The volume of the middle lobe of the right lung was 34.73% compared with the other lobes. The main lesions in each group were ground-glass density shadows, and the main lesions in group A were −570 ~ −470 HU density, accounting for 13.89%, followed by −470 ~ −370 HU, accounting for 11.07%. Only 3.22% and 4.75% of solid lesions with densities of 30 ~ 70 HU and −70 ~ 30 HU were found. Most of the lesions in group B were ground-glass density shadows, and the focal densities were mainly −570 ~ −470 HU, −470 ~ −370 HU, and −370 ~ −270 HU, accounting for 13.18%, 12.58%, and 12.52%, respectively. No statistical difference was noted between the proportion of lesions with a density of −570 ~ −470 HU and that of group A; however, the volume and proportion of other lesions with different densities were higher than those of group A, showing a trend that the higher the density of the lesions, the higher the proportion of group B was compared with group A. Conclusion: Larger infection volume, more lesion solid components, and multiple CT signs often indicate more severe lung inflammation, which easily progresses to severe disease. Quantitative measurement of chest CT based on deep learning is helpful for the prognosis assessment of COVID-19 and the early warning of severe outcome.

Keywords

• deep learning
• tomography
• x-ray computed
• coronavirus disease 2019