Comparison of temporal evolution of computed tomography imaging features in COVID-19 and influenza infections in a multicenter cohort study

2.1. Imaging

CT imaging protocols among the different centers were not strictly standardized, but imaging adhered to recommendations as follows: Patients were scanned in prone position, intubated patients were scanned in supine position. If CT angiography was required to evaluate for pulmonary embolism, patients were scanned in supine position with administration of intravenous iodinated contrast agent. Scanner settings were 100 kV, 50–100 mAs, depending on requirement for contrast agent or availability of automated dose modulation (e. g., automated tube voltage selection and automated tube current modulation). Reconstructions were made using iterative algorithms. Slice thickness was 5 mm or 2 mm for transverse and 5 mm for sagittal reconstructions. Imaging was performed as clinically indicated according to the referring physicians.

2.2. Image evaluation

Two radiologists (TF and YEB) with seven and eight years of experience in chest radiology, respectively, independently evaluated all CT examinations (axial and sagittal reconstructions), blinded for examination date, all clinical and microbiological information. Disagreements between both readers were processed as follows: Continuous variables: values were averaged; categorial variables: solved by consensus reading.

2.3. Evaluation of lobe involvement

Estimation of affected lung was performed as previously reported [11], [12] and was done lobe-wise (right upper lobe, middle lobe, right lower lobe, left upper lobe and left lower lobe) in percent (%) in each CT examination. Weighted mean of total lung involvement was calculated by use of previously described proportions [13].

2.4. Evaluation of main pattern

For each CT examination, the predominant pattern was chosen between the three following CT features: GGO (region of increased lung attenuation with vessels still visible), crazy paving (appearance of GGO with superimposed interlobular and intralobular septal thickening) or consolidation (homogeneous increased lung opacity with vessels not visible).

2.5. Evaluation of additional findings

Additional findings were evaluated as present or absent in each CT examination and were defined based on the most typical COVID-19 imaging features [14] and previously defined typical imaging features of influenza [3], [4], [5], [6], [7], [8], [15], [16]. Additional findings were: subpleural linear opacity, septal thickening, subpleural reticulation, air bronchogram, pleural thickening, bronchiectasis, bronchial wall thickening, tree-in-bud, pulmonary nodules, and vascular enlargement. Intrathoracic lymph node size (short-axis diameter), pleural effusion (if bilateral, both sides were summed up) and pericardial effusion were measured in millimeters. For evaluation of bronchial wall thickness (lobar, segmental and subsegmental), the T/D ratio (wall thickness (T) divided by total diameter of bronchus (D)) was measured. Values between 0.1 and 0.2 have been described as normal [17], [18]. In this study, a value above 0.2 was defined as bronchial wall thickening.

Nonspecific findings (neither systematically reported for COVID-19 nor influenza) were cavitation, halo sign, reverse halo sign and tracheal wall irregularities, that could reflect ulcers or pseudomembranes.

2.6. Clinical features

The following co-morbid features were recorded: Obesity, substance abuse, chronic obstructive pulmonary disease (COPD), asthma, neoplasm, hematological disease, hematopoietic stem cell transplantation (HSCT), solid organ transplantation (SOT), human immunodeficiency virus (HIV) infection, chronic renal failure, diabetes, cardiovascular disease, cerebrovascular disease, and pregnancy. The following medication prior to COVID-19 or influenza disease were recorded: Chronic steroid treatment (≥0.1 mg/kg/day prednisone equivalent) and immunosuppressive drugs. The following medication during the COVID-19 or influenza disease were recorded: Steroids and specific antiviral drugs. Additionally, bacterial respiratory co-infections were recorded according to assessment by infectious disease consultants and evaluated. The intubation status during the hospital stay was reviewed.

2.7. Data transformations

“Main pattern” was dichotomized into the three binary variables “consolidation”, “crazy paving”, and “ground glass opacity” being either present or absent. To simplify matters, multiple examinations per patients were reduced to one value for patients-wise comparisons. In case of involvement that was analyzed as a function of time, the method of least squares was used to compute the intercept and slope for each patient with more than one CT and the predicted involvement was calculated at the patient’s mean day after diagnosis. For all other variables, the mean was calculated for continuous variables and the maximum for binary variables (that were coded as 0 = absent and 1 = present) for each patient.

2.8. Statistics

Continuous variables with normal distribution were compared with a Student’s t-test for independent samples. Other continuous variables were compared with a Wilcoxon rank-sum test (with continuity correction) between COVID-19 and influenza patients. Categorical variables were compared with a Fisher’s exact test, intubation rates were compared with the chi-squared-test.

The intra-class correlation coefficient (ICC) was computed to assess the agreement between two radiologists rating the involvement of upper, lower, and total lung, left and right pleural effusion, lymph node size, and pericardial effusion in each of the 179 images. The ICC (two-way random-effects model for single rating) is reported. ICC is interpreted as poor for values below 0.50, as moderate for values between 0.50 and 0.75, as good for values between 0.75 and 0.90, and as excellent for values above 0.90 [19].

Cohen’s κ was computed to assess the agreement between the two radiologists rating various features as being either present or absent. The following classifications has been suggested to interpret the strength of the agreement based on the Cohen’s κ value [20]: Values > 0.4 were considered moderate, > 0.6 substantial and > 0.8 almost perfect agreement.

The association of diagnosis (influenza vs COVID-19) with combinations of involvement, GGO, air bronchogram, bronchiectasis, tree-in-bud, pulmonary nodules, pleural effusion, and pericardial effusion was further analyzed with tree-based models, which is a non-parametric procedure that works by partitioning the patients into subsets with similar response values using a set of splitting rules. The general procedure is as follows: Starting with the entire data set, a split is made at a specific cut point on one of the predictors chosen to maximize the difference in response between the resulting subsets. Each of the two subsets can be further subdivided using different predictors and/or different cut points. Splitting is associated with a test of significance for the difference in response between the two newly created subsets. Splitting continues as long as the resulting subsets are significantly different at the 5 % level (with or without correction for multiple testing). Analyses were mostly performed without correction for multiple testing and must therefore be regarded as exploratory.

2.9. Software

Most analyses were performed in the R programming language (version 4.0.2) (R Core Team, 2020). The package “psych” [21] was used to calculate ICC and Cohen’s κ. The package “tableone” [22] was used for descriptive statistics. The package “ggplot2″ [23] was used to plot the figures. The package “party” [24] was used to run the classification tree analysis. Comparison of lung involvement, main pattern and additional chest findings for COVID-19 patients and influenza patients with and without bacterial superinfection was done with Kruskal-Wallis Test (continuous variables) and Fisher’s exact test (categorical variables). Comparison of age was done with SPSS v. 25.0.0.2 (IBM Corp, Armok, NY) using Student’s t-test for independent samples.

3.1. Baseline characteristics

Patients’ mean age was 60.8 years (standard deviation (SD) 13.5 years, range 18.4–81.9 years), (COVID-19 patients 63.5 ± 9.5 years, range 35.2–80.7 years, influenza patients 57.6 ± 16.6 years, range 18.4–81.9 years). COVID-19 patients were significantly older (p = 0.04).

Overall, 34/96 (35.4 %), (12/52 (23.1 %) of COVID-19 patients and 22/44 (50.0 %) of influenza patients) were female. Males were more prevalent in the COVID-19 group (p = 0.01). Mortality was 21/52 (40.4 %) in COVID-19 patients and 10/44 (22.7 %) in influenza patients (p = 0.08). In the COVID-19 group, median time from disease specific symptom onset to the diagnosis was 3 days (IQR: 1–6 days), range − 3 to 14 days. In the influenza group, median time from disease specific symptom onset to the diagnosis was 4 days (IQR: 2–7 days), range 0–26 days.

3.2. Comparison of clinical data

Comparison of clinical data is given in Table 1. Obesity was more common among COVID-19 patients (p < 0.001), whereas substance abuse (p = 0.003), COPD (0.019) and chronic renal failure (0.022) were more common among influenza patients. Prior chronic treatment with steroids was more common among influenza patients (p = 0.046). COVID-19 patients were more commonly treated with steroids (p < 0.001), whereas influenza patients were more commonly treated with antivirals, which was oseltamivir in all cases (p < 0.001). Of patients with COVID-19, 23 (44 %) were intubated, and 19 (43 %) of patients with influenza were intubated, differences were not significant (p = 1).

3.3. Interrater reliability

Interrater reliability was excellent for total lung, upper and lower lung involvements, each with ICC > 0.9. It was good for left and right pleural effusion as well as for pericardial effusion, and moderate for lymph node size. Cohen’s κ could not be computed for reverse halo sign, where for 178 images, both readers rated “absent” and for one image only, one reader rated “present”. For most other imaging features, the agreement was substantial to near perfect.

3.4. Comparison of lung involvement

There was greater lung involvement for COVID-19 patients with median differences of 21.5 % for total lung involvement (p < 0.001), of 20.6 % for upper lobe involvement (p < 0.001) and of 7.4 % for lower lobe involvement (p = 0.005) (Table 2).

3.5. Comparison of lung involvement over time

Time-dependent evaluation of lung involvement revealed a continuous increase over time for total lung, upper and lower lobe involvement in COVID-19 patients, whereas lung involvement increased in influenza patients until the end of the second week with a subsequent decrease (Fig. 1).

3.6. Comparison of main pattern

In an overall comparison of observed main patterns (consolidation, crazy paving and GGO), GGO was the predominant pattern in COVID-19 patients (Table 2). While GGO was the most common main pattern in COVID-19 patients (71.2 % vs. 38.6 % in influenza, p = 0.002), consolidation was the most common main pattern in patients with influenza (50.0 % vs. 30.8 % in COVID-19, p = 0.06). Presence of bacterial superinfection did not alter these differences between COVID-19 and influenza (data not shown).

3.7. Temporal comparison of predominant main pattern

For the majority of COVID-19 cases, GGO was changing from present to absent or was not changing over time. Changes from absent to present later in the course of the disease were uncommon. For the majority of influenza patients, GGO was not changing over the whole period of observation. The probability of GGO was lower for influenza than for COVID-19 patients, especially in earlier disease (≤ day seven). The probability seemed to be declining over time for both patient groups (Fig. 2a). Although crazy paving could change within patients from present to absent and vice versa, for more than half of all patients the predominant crazy paving pattern did not change over time. Crazy paving as main pattern was slightly increasing over time both for COVID-19 and for influenza patients (Fig. 2b). The probability for consolidation as main pattern was increasing over time for COVID-19 patients and decreasing for influenza patients (Fig. 2c).

3.8. Comparison of additional findings

Comparison of additional findings between COVID-19 and influenza patients indicated that air bronchogram was present more often in COVID-19 (63.5 %) than influenza patients (38.6 %, p = 0.024, Table 3). Conversely, tree-in-bud (p < 0.001) and pulmonary nodules (p = 0.04) were present more often in influenza (40.9 % and 18.2 %, respectively) than COVID-19 patients (5.8 % and 3.8 %, respectively). Moreover, the size of pleural and pericardial effusion was larger in influenza than COVID-19 patients (Table 3). Taking the presence of bacterial superinfection into account slightly changed these results (Table 4). Air bronchogram and pleural effusion were not different any more (p = 0.08 and p = 0.08, respectively) while tracheal wall irregularities were present only in those without bacterial superinfection and more commonly in COVID-19 (16.7 % in COVID-19 patients vs. 4.8 % in influenza patients, p = 0.02).

3.9. Outcome prediction

For influenza and COVID-19 patients, none of the evaluated variables seemed to be strongly associated with death (p-value >0.05 in bivariate analyses for all, including maximum lung involvement).

3.10. Classification tree analysis

A classification tree for prediction of COVID-19 was constructed for the entire observation time with Bonferroni correction for multiple testing (Fig. 3). Degree of lung involvement, absence of tree-in-bud, and smaller pericardial effusion were significant predictors for COVID-19 (compared to influenza).

Four different classification trees for the prediction of COVID-19 were developed for different time points (Fig. 4a-d). For the first time bin, i.e., ≤ 0 days after diagnosis, only pleural effusion was retained: Larger volumes of pleural effusion were associated with influenza (Fig. 4a). For the second time bin, i.e., 1–7 days after diagnosis, pericardial effusion and GGO were retained in the partitioning: A lack of pericardial effusion at 1–7 days after diagnosis was a significant predictor for COVID-19, especially if in combination with GGO (Fig. 4b). For the third time bin, i.e., 8–14 days after diagnosis, pleural effusion and lung involvement were retained in the partitioning: Pleural effusion ≤ 9.5 cm (both sides summed up) at 8–14 days after diagnosis was a significant predictor for COVID-19, especially if in combination with larger total lung involvement (Fig. 4c). For the fourth time bin, i.e., > 14 days after diagnosis, pleural effusion and lung involvement were retained in the partitioning:Pleural effusion ≤ 9.5 cm (both sides summed up) at > 14 days after diagnosis was a significant predictor for COVID-19, especially if in combination with larger total lung involvement (Fig. 4d). Examples of imaging features associated with influenza are given in Fig. 5.