This retrospective study was approved by the Institutional Review Board of the Pusan National University Hospital (2004-034-090), which waived the requirement for informed consent. This study used the open data repository for COVID-19, namely, the Korean Imaging Cohort for COVID-19 (KICC-19), which was constructed by the Korean Society of Thoracic Radiology (KSTR) during July 2020.89 From February through May 2020, 522 adult patients (≥ 18 years old) with real-time reverse transcriptase-polymerase chain reaction (RT-PCR) results and positivity for coronavirus by nasal and oropharyngeal swab testing were enrolled in KICC-19. Other respiratory pathogens capable of causing symptoms were excluded by RT-PCR and sputum culture. Of these 522 patients, 271 underwent chest CT within 10 days of diagnosis and constituted the study cohort. Mean time between symptom onset and chest CT examination was 5.5 ± 2.7 days (range, 0–10 days).

Demographics, clinical symptoms, underlying diseases, initial laboratory findings including lymphocyte counts, C-reactive protein (CRP), procalcitonin, type of treatment, duration of hospital stay, clinical severity, and clinical outcome (cure or death) of the 271 study subjects were evaluated. Lymphocytopenia was defined as fewer than 1,000 lymphocytes per microliter of blood.1011 The thresholds for CRP and procalcitonin elevation were 0.5 mg/dl and 0.05 ng/ml, respectively.1213 The 271 patients were divided into 2 groups, that is, a severe group of patients who needed additional oxygen therapy or mechanical ventilation, and a mild group of patients who did not.

All CT examinations were performed using multi-detector CT scanners with more than 16 channels. Scans were obtained from lung apices to lung bases. Intravenous contrast medium was injected in only 9 patients (3%). Of the 271 scans, 133 (49%) and 130 (48%) were obtained using a slice thickness of ≤ 1.5 mm or between 1.5 and 3 mm, respectively; only 8 CT scans (3%) were constructed using a slice thickness of > 3 mm.

All images were anonymized locally and transferred to a cloud-based web interface (AiCRO System, Asan Image Metrics (AIM), Strategy & Development Unit, Clinical Trial Center, Asan Medical Center, Korea) for analysis. CT data were also transferred to a post-processing workstation (syngo.via, Siemens) for further analysis by an investigator with 18 years' experience in thoracic imaging.

Two radiologists with 9 and 13 years of chest CT experience unaware of patient clinical severity, reviewed all images and reached conclusions by consensus. Only the first CT examination was analyzed if a patient underwent multiple CT examinations within 10 days of COVID-19 diagnosis. Times between symptom onsets and CT examinations were recorded. CT findings were evaluated for the presence, distribution, and extents of parenchymal abnormalities including GGO, consolidation, crazy-paving appearance, reversed halo sign, interlobular and intralobular interstitial thickenings, reticulation, subpleural lines, traction bronchiectasis, air-bronchograms, large nodules, and centrilobular nodules. A crazy-paving appearance was defined as thickened interlobular septa and intralobular lines superimposed on a background of GGO, and the reversed halo sign was defined as a focal rounded area of GGO surrounded by a near or complete ring of consolidation.14 The presences of vascular enlargement, pleural effusion, hilar, and mediastinal lymph node enlargement were also evaluated. An enlarged lymph node was defined as one with a short-axis diameter of > 10 mm. In addition, the presence of vascular enlargement was evaluated and defined as a vessel diameter larger than that expected based on its location in the vascular tree, than in an adjacent portion of non-diseased lung, or in a comparable region of non-diseased contralateral lung or focal dilation or non-tapering of vessels as they coursed toward lung periphery. Vessel enlargement within areas of parenchymal opacity, outside of opacities, and dilated distal subsegmental vessels touching pleura or fissure were recorded.15

Total extents of all parenchymal lesions were assessed using a visual scoring system and artificial intelligence (AI) software (CT Pneumonia Analysis 2.1.2, syngo,via FRONTIER, Siemens). With a visual CT scoring system, each lung lobe was visually scored from 0 to 5 as: 0, no involvement; 1, < 5% involvement; 2, 5%–24%; 3, 25%–49%; 4, 50%–74%; 5, ≥ 75%. Visual CT scores were the sum of the individual lobar score and ranged from 0 to 25. With AI software for the automatic quantification of COVID-19, the total extent of lung parenchymal abnormality was calculated. Data from 245 patients were analyzed in automatic measurement; 26 patients were excluded due to data non-availability.

Laterality (unilateral and bilateral) and distributions of lung parenchymal abnormalities in the transverse (central, peripheral, peribronchovascular, and random) and longitudinal planes (upper zone, middle zone, lower zone, and random) were evaluated.

Parenchymal abnormalities were then classified into; bronchopneumonia, OP, or DAD patterns. When parenchymal abnormalities contained areas of consolidation, GGO, large nodules, centrilobular nodules, and bronchial wall thickening, they were considered to exhibit a bronchopneumonia pattern. OP patterns were considered to show lung abnormalities consisting of consolidation and GGO with subpleural or peribronchovascular predominance, and DAD patterns were considered to demonstrate patchy or extensive airspace consolidation or GGO without zonal predominance.

Statistical analysis was performed using SPSS Ver. 21.0 software (SPSS, IBM SPSS Statistics, Armonk, NY, USA). Results are expressed as means (SDs) for continuous variables and as numbers of individuals and percentages for the categorical variables.

The Mann-Whitney U test, the χ² test, and Fisher's exact test were used to determine the significances of differences between clinical and CT features of the severe and mild groups. Factors associated with poor prognosis were identified by univariate and multivariate logistic regression analysis. The comparison of two prediction models, which included visual CT scores and AI measurement, respectively, was performed with a receiver operating characteristic (ROC) analysis. Statistical significance was accepted for P values of < 0.050 throughout.

The present study protocol was approved by the Institutional Review Board of the Pusan National University Hospital (2004-034-090), which waived the requirement for informed consent.

The demographics and clinical data of the 271 patients, all of whom performed a chest CT scan, are presented in Table 1. There were 116 men and 155 women of mean age 52.9 ± 19.5 years (range, 15–97 years). One hundred and four (38.4%) had underlying disease: hypertension (n = 71), diabetes (n = 46), cardiovascular disease (n = 13), or cancer (n = 27). One hundred and fifty-seven were febrile and 131 had respiratory symptoms such as a cough with or without sputum production. One hundred and twenty-two patients were treated with lopinavir/ritonavir and 83 with hydroxychloroquine. Mean hospital stay was 25.2 ± 15.6 days (range, 0–94 days). Although most patients had cured, ten patients (3.7%) died.

Eighty-seven (32.1%) had lymphocytopenia and 154 (56.8%) elevated CRP at admission. An elevated procalcitonin level was observed in 76 (28%) patients. Fifty-five (20.3%) needed additional oxygen therapy or mechanical ventilation, and 216 did not (Table 2). Of the 55 patients in the severe group, 10 died of respiratory failure. Regarding demographics, only age was significantly different in severe and mild groups (P < 0.001). Lymphocytopenia, elevated CRP, and procalcitonin were more commonly observed in the severe group (Ps < 0.050)

Of the 271 patients, 45 (16.6%) had normal chest CT findings. CT findings of COVID-19 pneumonia in the other 226 (83.4%) patients are presented in Table 3. The most common CT finding among these 226 patients was GGO (98%, 221 patients), followed by consolidation (53%, 120 patients), intralobular interstitial thickening (34%, 77 patients), and a crazy-paving appearance (25%, 56 patients). Other findings (in decreasing order of frequency) were air bronchogram, (19%, 42 patients), reversed halo (8%, 17 patients), reticulation (5%, 12 patients), traction bronchiectasis (5%, 11 patients), pleural effusion (4%, 8 patients), centrilobular nodules (3%, 7 patients), subpleural line (3%, 7 patients), and interlobular septal thickening (3%, 6 patients). Only twelve patients (5%) had vascular enlargement and the majority (11 patients) showed vessel enlargement within areas of parenchymal opacity; vessel enlargement outside opacity was observed in only one patient.

The parenchymal abnormalities were bilateral in 169 patients (75%) and unilateral in 57 patients (25%). In the longitudinal plane, lung abnormalities showed mainly lower lung involvement (50%, 112 patients) and random involvement (46%, 103 patients). In the transverse plane, the lung abnormalities showed mainly peribronchovascular predominance (57%, 129 patients) and all 226 exhibited a peripheral or peribronchovascular distribution. The mean visual CT score of parenchymal abnormalities was 7.51 ± 5.56 (range, 1–23) and mean total extent by automatic measurement was 10.5 ± 16.2 (range, 0.01%–73.6%).

The CT findings were classified as OP (93%, 210 patients), DAD (4%, 9 patients), or bronchopneumonia (3%, 7 patients) (Figs. 1 and 2). Table 4 summarizes CT findings for the severe and mild groups. GGO, consolidation, intralobular and interlobular interstitial thickening, crazy-paving appearance, reticulation, traction bronchiectasis, air-bronchogram, mediastinal and hilar lymph node enlargement were most frequently observed in the severe group (all P < 0.050). Bilateral parenchymal abnormalities were more common in the severe group (P < 0.001). Random distribution in the longitudinal plane and peripheral and peribronchovascular distribution in the transverse plane were more frequently observed in the severe group (both P < 0.001). Mean visual CT score and total extent of automatic measurement were higher in the severe group (11.8 vs 4.9, P < 0.001 and 27.5% vs 5.9%, P < 0.001, respectively). All patients with DAD patterns were in the severe group. Of the 10 patients that died, 3 showed a DAD pattern and the remaining 7 showed an OP pattern (Fig. 2).

Univariate analyses showed older age, elevated CRP and procalcitonin levels, and CT findings including GGO, lymph node enlargement, bilateral involvement of parenchymal abnormalities, visual CT scores, and total extent by automatic measurement of lung parenchymal abnormalities were significant prognostic parameters. Multivariate logistic analysis with two models, which included visual CT scores and AI measurement, respectively, showed elevated CRP, elevated procalcitonin, visual CT scores, and AI measurement of lung parenchymal abnormalities were independently associated with a poorer prognosis for patients with COVID-19 pneumonia (Table 5). There was no significant difference of two models with results of area under the ROC curves: 0.8587 vs. 0.8645, P = 0.632 (Fig. 3).