基于深度学习的新型冠状病毒肺炎转归胸部CT评价

目的 分析基于深度学习的新型冠状病毒肺炎（COVID-19）不同临床转归患者胸部CT的差异，以提高对COVID-19转归的影像认识。 方法 回顾性分析2020年1月25日至3月29日来自内蒙古自治区COVID-19病例库的42例COVID-19患者（临床分型为普通型）的胸部CT资料，其中，男性20例、女性22例，年龄17个月~86岁（48.74±17.18）岁]。根据是否转为重症（重型或危重型）将患者分为未转为重症的A组（n=29）和转为重症的B组（n=13），比较2组患者年龄、性别及基于深度学习的胸部CT表现，评价感染肺叶病灶分布，累及肺叶侧别、数目，感染肺叶病灶体积，密度（CT值）和感染肺叶病灶体积占比等资料的差异。计量资料的比较采用两独立样本t检验、Mann-Whitney U非参数检验；计数资料的比较采用卡方检验或Fisher's确切概率法。 结果 2组患者的性别差异无统计学意义（χ2=0.016，P=1.000）。B组患者的年龄高于A组（65.62±11.24）岁对（41.17±13.66）岁 ]，且差异有统计学意义（t=5.64，P<0.001）。B组患者感染肺叶数以及各肺叶感染体积占比、总感染体积占比均高于A组，且差异有统计学意义（Z=2.505~3.605，均P<0.05）。2组患者肺部总体积差异无统计学意义（Z=1.456，P=0.146），B组患者各肺叶感染体积及双肺总感染体积均高于A组，且差异有统计学意义（Z=2.301~3.254，均P<0.05）；B组患者在各CT阈值范围内的肺部感染体积占比均高于A组，且差异有统计学意义（Z=3.115~3.578，均P<0.05）。胸部CT和人工智能病灶识别图的图像结果显示，病灶均以磨玻璃密度影、实变为主，双肺下叶感染较多，右肺中叶较少。 结论 转为重症的COVID-19患者的胸部CT明显有别于未转为重症的患者。基于深度学习的人工智能可尽早评估有重症转归倾向的患者，有助于COVID-19重症率的控制。

Chest CT evaluation of COVID-19 outcome based on deep learning

Objective To analyze variations in the chest CT of different clinical outcomes of corona virus disease 2019 (COVID-19) based on deep learning and improve the understanding on COVID-19 imaging. Methods The chest CT of 42 cases (the clinical classification was common type) of COVID-19 in Inner Mongolia Autonomous Region collected from January 25, 2020 to March 29, 2020 were examined. The cases included 20 males and 22 females, with ages ranging from 17 months to 86 (48.74±17.18) years. The patients were divided into group A, which included those did not progress to severe disease (n=29), and group B, which included those who progressed to severe/critical disease (n=13). Differences in age, gender, lesion distribution, sides, number, volume, density (CT value), and proportion of lesion volume as detected by chest CT were compared between the two groups by deep learning. Two independent samples t test and Mann-Whitney U nonparametric tests were used to compare measurement data, and the χ2 and Fisher's exact tests were used to compare count data. Results No statistical difference in gender was noted between the two groups (χ2=0.016, P=1.000). The mean age of group B was higher than that of group A (65.62±11.24 years vs. 41.17±13.66 years), and a statistical difference was observed in each group (t=5.64, P<0.001). The number of infected pulmonary lobes, proportion of infection volume in each pulmonary lobe, and proportion of total infection volume were higher in group B than in group A, and a statistical difference was noted in each group (Z=2.505?3.605, all P<0.05). No statistical difference in total lung volume between the two groups was observed (Z=1.456, P=0.146). The size of infection in each lobe and the total volume of infection in both lungs in group B were greater than those in group A, and a statistical difference was found in each group (Z=2.301?3.254, all P<0.05). The proportion of lesions in group B within the threshold range of all CT values was higher than that in group A, and a statistical difference was observed in each group (Z=3.115?3.578, all P<0.05). The results of chest CT and artificial intelligence lesion recognition mapping revealed that lesions in serious cases are mainly characterized with ground glass opacity and consolidation. Moreover, the lesions frequently involved the lower lobes of the lungs and less commonly affected the middle lobe of the right lung. Conclusions The chest CT of patients with COVID-19 who progressed to severe disease and those who did not showed significant differences. Artificial intelligence based on deep learning can assess patients with a tendency to progress to severe/critical disease early and contribute to the improved management of severe COVID-19.