From the Department of Radiology, Hanyang University Medical Center, Hanyang University College of Medicine, Seoul, South Korea.
CHESS Center, The First Hospital of Lanzhou University, Lanzhou, China.
J Comput Assist Tomogr. 2022;46(3):413-422. doi: 10.1097/RCT.0000000000001303. Epub 2022 Apr 8.
We aimed to develop and validate the automatic quantification of coronavirus disease 2019 (COVID-19) pneumonia on computed tomography (CT) images.
This retrospective study included 176 chest CT scans of 131 COVID-19 patients from 14 Korean and Chinese institutions from January 23 to March 15, 2020. Two experienced radiologists semiautomatically drew pneumonia masks on CT images to develop the 2D U-Net for segmenting pneumonia. External validation was performed using Japanese (n = 101), Italian (n = 99), Radiopaedia (n = 9), and Chinese data sets (n = 10). The primary measures for the system's performance were correlation coefficients for extent (%) and weight (g) of pneumonia in comparison with visual CT scores or human-derived segmentation. Multivariable logistic regression analyses were performed to evaluate the association of the extent and weight with symptoms in the Japanese data set and composite outcome (respiratory failure and death) in the Spanish data set (n = 115).
In the internal test data set, the intraclass correlation coefficients between U-Net outputs and references for the extent and weight were 0.990 and 0.993. In the Japanese data set, the Pearson correlation coefficients between U-Net outputs and visual CT scores were 0.908 and 0.899. In the other external data sets, intraclass correlation coefficients were between 0.949-0.965 (extent) and between 0.978-0.993 (weight). Extent and weight in the top quartile were independently associated with symptoms (odds ratio, 5.523 and 10.561; P = 0.041 and 0.016) and the composite outcome (odds ratio, 9.365 and 7.085; P = 0.021 and P = 0.035).
Automatically quantified CT extent and weight of COVID-19 pneumonia were well correlated with human-derived references and independently associated with symptoms and prognosis in multinational external data sets.
我们旨在开发和验证用于计算机断层扫描（CT）图像的 2019 年冠状病毒病（COVID-19）肺炎的自动定量方法。
本回顾性研究纳入了 2020 年 1 月 23 日至 3 月 15 日来自 14 家韩国和中国机构的 131 例 COVID-19 患者的 176 例胸部 CT 扫描。两名有经验的放射科医生在 CT 图像上半自动地绘制肺炎口罩，以开发用于分割肺炎的 2D U-Net。使用日本（n=101）、意大利（n=99）、Radiopaedia（n=9）和中国数据集（n=10）进行外部验证。系统性能的主要衡量标准是与视觉 CT 评分或人类分割相比，肺炎的程度（%）和重量（g）的相关系数。多变量逻辑回归分析用于评估日本数据集中症状与程度和重量的关联，以及西班牙数据集中复合结局（呼吸衰竭和死亡）（n=115）。
在内部测试数据集，U-Net 输出与参考之间的内部一致性系数为 0.990 和 0.993，用于程度和重量。在日本数据集中，U-Net 输出与视觉 CT 评分之间的皮尔逊相关系数为 0.908 和 0.899。在其他外部数据集，内部一致性系数在 0.949-0.965 之间（程度）和 0.978-0.993（重量）之间。四分位区间内的程度和重量与症状独立相关（比值比，5.523 和 10.561；P=0.041 和 0.016）和复合结局（比值比，9.365 和 7.085；P=0.021 和 P=0.035）。
COVID-19 肺炎的自动定量 CT 程度和重量与人类参考高度相关，并且在多国外部数据集中与症状和预后独立相关。
