Automated Quantitative Lung CT Improves Prognostication in Non-ICU COVID-19 Patients beyond Conventional Biomarkers of Disease

(1) Background: COVID-19 continues to represent a worrying pandemic. Despite the high percentage of non-severe illness, a wide clinical variability is often reported in real-world practice. Accurate predictors of disease aggressiveness, however, are still lacking. The purpose of our study was to evaluate the impact of quantitative analysis of lung computed tomography (CT) on non-intensive care unit (ICU) COVID-19 patients’ prognostication; (2) Methods: Our historical prospective study included fifty-five COVID-19 patients consecutively submitted to unenhanced lung CT. Primary outcomes were recorded during hospitalization, including composite ICU admission for the need of mechanical ventilation and/or death occurrence. CT examinations were retrospectively evaluated to automatically calculate differently aerated lung tissues (i.e., overinflated, well-aerated, poorly aerated, and non-aerated tissue). Scores based on the percentage of lung weight and volume were also calculated; (3) Results: Patients who reported disease progression showed lower total lung volume. Inflammatory indices correlated with indices of respiratory failure and high-density areas. Moreover, non-aerated and poorly aerated lung tissue resulted significantly higher in patients with disease progression. Notably, non-aerated lung tissue was independently associated with disease progression (HR: 1.02; p-value: 0.046). When different predictive models including clinical, laboratoristic, and CT findings were analyzed, the best predictive validity was reached by the model that included non-aerated tissue (C-index: 0.97; p-value: 0.0001); (4) Conclusions: Quantitative lung CT offers wide advantages in COVID-19 disease stratification. Non-aerated lung tissue is more likely to occur with severe inflammation status, turning out to be a strong predictor for disease aggressiveness; therefore, it should be included in the predictive model of COVID-19 patients.


Introduction
More than a year after the first cases in Wuhan, COVID-19 continues to represent a worrying pandemic considering numbers and real-world variability. Despite a broader representation of asymptomatic or paucisymptomatic cases, COVID-19 can lead to severe illness in up to 14% of patients. Moreover, 5% become critical with 49% of mortality in some case series [1]. Thus, identifying patients with potential for severe or critical illness should be considered of primary importance.
Several studies have tried to identify potential diagnostic and predictive models in the COVID-19 approach, unfortunately often reporting a high risk of bias [2][3][4]. Therefore, the clinical validity of different predictors remains undetermined.
Diagnostic imaging also has an uncertain role [13][14][15][16][17][18][19]. Evidence of this can be found in the current World Health Organization (WHO) recommendations on chest imaging in the diagnosis and management of COVID-19, which are based only on low-level evidence or expert recommendations, meaning without significant support from the literature [20].
In this scenario, WHO recommendations suggest chest imaging mainly as support to clinical and laboratory assessment to decide the most adequate management of COVID-19 patients, although nonquantitative parameters to be considered are provided [20].
Among different imaging tools, computed tomography (CT) can offer significant advantages in risk stratification.
Lung CT finds encouraging results also in describing the disease, quantifying the burden of disease and as a disease-progression predictor, beyond conventional clinical and respiratory parameters used in real-world practice.
Different analysis methods have been applied in clinical practice, most of them based on qualitative (i.e., description of lesion characteristics including ground glass opacities or consolidation) or semiquantitative analysis (i.e., severity score or overall score based on percentage of different lobe or segment involvement).
More recently, lesion volume measurement through the identification of different densitometric thresholds, as a quantitative method, has become a promising approach to COVID-19 [38,39].
In a recent paper of Chiumello et al., quantitative CT patterns resulted advantageous also in describing COVID-19 as a nonconventional subset of acute respiratory distress syndrome (ARDS) when compared to lung physiology [39].
Going back to these premises, the purpose of our study was to evaluate the impact of quantitative lung CT including volumetric and tissue weight assessment of differently aerated areas of lung on non-intensive care unit (ICU) COVID-19 patients' prognostication. Figure 1. Automatic lung volume detection. Two different views from a volume-rendered reconstruction were obtained from the automatic detection of lung tissue (img (a,b)). Pink and yellow arrows in img (c) highlight also high-density lesion within lung tissue, in a coronal view. In images (d,e), magnified, volume-rendered reconstruction and coronal CT scan view (respectively) enhance the high capability of CT and quantitative reconstruction to detect also millimetric lesion. Automatic lung volume detection. Two different views from a volume-rendered reconstruction were obtained from the automatic detection of lung tissue (img (a,b)). Pink and yellow arrows in img (c) highlight also high-density lesion within lung tissue, in a coronal view. In images (d,e), magnified, volume-rendered reconstruction and coronal CT scan view (respectively) enhance the high capability of CT and quantitative reconstruction to detect also millimetric lesion.
Different studies also showed the effectiveness of quantitative lung assessment in prediction of disease progression, limiting the analysis to the volumetric quantification of lesion.
However, lung tissue weight can be derived from quantitative analysis through volume and mean density assessment, thus allowing a specific differentiation between areas with different gas/tissue ratios, which can turn critically in the clinical work-up of a symptomatic COVID-19 patient [39,[45][46][47][48][49][50].
Going back to these premises, the purpose of our study was to evaluate the impact of quantitative lung CT including volumetric and tissue weight assessment of differently aerated areas of lung on non-intensive care unit (ICU) COVID-19 patients' prognostication.

Materials and Methods
This study was carried out after the approval of our university's Internal Review Board committee. This is a retrospective assessment of prospectively followed-up patients (historical prospective study).
All patients had documented COVID-19 (i.e., positive reverse transcriptase polymerase chain reaction (RT-PCR) on nasal or pharyngeal swab). Hospitalized non-ICU COVID-19 patients submitted to lung CT for an adequate work-up at admission were included. Figure 2 shows the flowchart of our study.

Materials and Methods
This study was carried out after the approval of our university's Internal Review Board committee. This is a retrospective assessment of prospectively followed-up patients (historical prospective study).
All patients had documented COVID-19 (i.e., positive reverse transcriptase polymerase chain reaction (RT-PCR) on nasal or pharyngeal swab). Hospitalized non-ICU COVID-19 patients submitted to lung CT for an adequate work-up at admission were included. Figure 2 shows the flowchart of our study.

Exam Protocol
Unenhanced CT examinations were performed with a Canon Aquilion One (320 rows detectors, 0.5 mm collimation; Canon Medical Systems, Otawara, Japan) (120 kv, ADE; mean dose < 5 mSv). All examinations were acquired at room air (RA). Whole-lung CT was performed under static conditions during an end-inspiratory hold. CT volumes were reconstructed in both lung and mediastinal windows (W: 1600 L: −550 and W: 40A0 L: 40, respectively).

Postprocessing Analysis
Postprocessing analysis was performed with dedicated software (CT Lung Density Analysis, Vitrea Advance Visualization, Canon) [51]. Notably, an automated segmentation of lung tissues with quantifiable controls and renderings is performed by the software. The Lung Density Analysis Tool segments the airways (including the trachea, main bronchi, and some larger bronchioles) and vascular structures, for both left and right lung. The contours of all the segmented structure are highlighted with different colors, and the possibility to edit contours for corrections is enabled.

Exam Protocol
Unenhanced CT examinations were performed with a Canon Aquilion One (320 rows detectors, 0.5 mm collimation; Canon Medical Systems, Otawara, Japan) (120 kv, ADE; mean dose < 5 mSv). All examinations were acquired at room air (RA). Whole-lung CT was performed under static conditions during an end-inspiratory hold. CT volumes were reconstructed in both lung and mediastinal windows (W: 1600 L: −550 and W: 40A0 L: 40, respectively).

Postprocessing Analysis
Postprocessing analysis was performed with dedicated software (CT Lung Density Analysis, Vitrea Advance Visualization, Canon) [51]. Notably, an automated segmentation of lung tissues with quantifiable controls and renderings is performed by the software. The Lung Density Analysis Tool segments the airways (including the trachea, main bronchi, and some larger bronchioles) and vascular structures, for both left and right lung. The contours of all the segmented structure are highlighted with different colors, and the possibility to edit contours for corrections is enabled.
The total lung tissue volume is computed using all segmented lung voxels, including unclassified voxels, that is, lung voxels outside the defined HU ranges.
The analysis was performed on total volume to avoid partial volume artifact. Calculation of weight of differently aerated areas was made through volumes and mean density (lung weight = lung volume × (mean density + 1000)/1000) [45][46][47].
Moreover, in a subset of ten examinations, lung volumes were compared with data obtained via an AI-based software [30]. The total lung tissue volume is computed using all segmented lung voxels, including unclassified voxels, that is, lung voxels outside the defined HU ranges.
The analysis was performed on total volume to avoid partial volume artifact. Calculation of weight of differently aerated areas was made through volumes and mean density (lung weight = lung volume × (mean density + 1000)/1000) [45][46][47].
Moreover, in a subset of ten examinations, lung volumes were compared with data obtained via an AI-based software [30].

Clinical Follow-Up Study
All patients were clinically followed up during their hospitalization in the Infectious Disease Clinic of our hospital. Symptoms, clinical risk factors, laboratory, and respiratory data were collected at admission and during hospitalization. Primary outcomes were composite ICU admission for the need of mechanical ventilation and/or death occurrence. Secondary outcomes considered death occurrence.

Statistical Analysis
Descriptive variables are presented as mean and correspondent confidential intervals or as percentages (frequencies). The Shapiro-Wilk (SW) test was used to evaluate data distribution. The distributional assumption for parametric analysis was fulfilled according to the SW test. A t-test was used for normal variables comparison; a chi-squared test was used with nominal (dichotomic) variables. Cox regression analysis was used to test the predictive validity of quantitative CT parameters. Variance inflation factor (VIF) was considered for evaluating multicollinearity. Kaplan-Maier was tested for qualitatively evaluating outcomes fitting for patients categorized for the median of the best predictor.
Differences between inter-software volume detection were tested with an independentsample t-test. Bland-Altman analysis with 95% limits of agreement was used for intersoftware agreement in assessment of total lung volumes and high-density volumes; differences are plotted as percentage.
To investigate the added value of quantitative parameters to predict disease progression, the following 4 models were used: Model I (only clinical characteristics), Model II (Model I + respiratory and laboratoristic characteristics), Model III (Model II + percentage of non-aerated lung volume), and Model IV (Model II + percentage of non-aerated tissue-weight). As a measure of discrimination, we calculated the area under the receiveroperating characteristics curve (C-index) with 95% confidence intervals (CIs) for diagnosis of significant disease progression. The added value of Models II, III, and IV beyond the basic model was quantified by the change in the C-index.
An alpha error of 5% was used as a threshold of significance, conventionally considered as an acceptable threshold for a conditional probability of 5% to experience a type I error. All statistical analyses were performed with SPSS (IBM Corp. Released 2016. IBM SPSS Statistics for Mac, Version 26.0. Armonk, NY, USA: IBM Corp).

Patient Population
Fifty-five patients (mean age 61 ± 14 years; 37 males) met inclusion criteria, i.e., patients consecutively admitted to the Infectious Disease Clinic of our hospital with (i) an RT-PCR-based diagnosis of COVID-19; (ii) respiratory failure in RA that did not require mechanical ventilation; (iii) CT with typical COVID-like pattern and bilateral infiltrations.
Patients with disease progression showed a higher percentage of obesity, hypertension, diabetes, and chronic obstructive pulmonary disease (COPD). Moreover, they were more often dyspneic. Fever was the most reported clinical sign (52 out of 55 patients, 95%, reported fever).
Antiviral drugs were administered to almost all patients (52 out of 55, 95%: p-value 0.006 in patients categorized for primary outcomes).

Laboratory and Respiratory Findings
Among different laboratory findings, neutrophile-to-leucocyte ratio (NLR), lactate dehydrogenase (LDH), and D-Dimer were significantly higher in patients with disease progression. Among respiratory data, SpO 2 and PF resulted significantly different between the two groups ( Table 1).
Bland-Altman plots of the agreement between different measure for total and highdensity volumes are shown in Figure 4.

Laboratory and Respiratory Findings
Among different laboratory findings, neutrophile-to-leucocyte ratio (NLR), lac dehydrogenase (LDH), and D-Dimer were significantly higher in patients with dise progression. Among respiratory data, SpO2 and PF resulted significantly different tween the two groups (Table 1).
Bland-Altman plots of the agreement between different measure for total and h density volumes are shown in Figure 4.

Quantitative Lung CT: Lung Parameter
Patients who reported disease progression showed lower total lung volume des similar weights. Moreover, non-aerated and poorly aerated lung tissue resulted sign cantly higher in patients with disease progression (p-value 0.003 and 0.011, respective well-aerated tissue resulted lower (p-value 0.011).
Conversely, overinflated tissue was similar between the two groups. Scores of non-aerated tissue expressed as a percentage of total weight (NAw%) percentage of total volume (NAv%) were also calculated, both of which resulted highe patients with disease progression (0.001 and 0.003, respectively). Similarly, high-den tissue, including both non-aerated and poorly aerated tissue, resulted higher in patie with disease progression when calculated as weight and volume percentage (HDw% HDv%, p-value 0.001 and 0.002, respectively) ( Table 1).

Comparison between CT Parameters and Laboratory/Respiratory Findings
Good correlation resulted when laboratory findings were compared with CT par eters, except for overinflated tissue.
Among respiratory findings, PaO2 and SpO2 did not correlate with CT paramet conversely, PF correlated with CT parameters, except for overinflated tissue (Table 2) D-Dimer and PF showed an inverse correlation (r −0.46, p-value 0.001).

Quantitative Lung CT: Lung Parameter
Patients who reported disease progression showed lower total lung volume despite similar weights. Moreover, non-aerated and poorly aerated lung tissue resulted significantly higher in patients with disease progression (p-value 0.003 and 0.011, respectively); well-aerated tissue resulted lower (p-value 0.011).
Conversely, overinflated tissue was similar between the two groups. Scores of non-aerated tissue expressed as a percentage of total weight (NAw%) and percentage of total volume (NAv%) were also calculated, both of which resulted higher in patients with disease progression (0.001 and 0.003, respectively). Similarly, high-density tissue, including both non-aerated and poorly aerated tissue, resulted higher in patients with disease progression when calculated as weight and volume percentage (HDw% and HDv%, p-value 0.001 and 0.002, respectively) ( Table 1).

Comparison between CT Parameters and Laboratory/Respiratory Findings
Good correlation resulted when laboratory findings were compared with CT parameters, except for overinflated tissue.
Among respiratory findings, PaO 2 and SpO 2 did not correlate with CT parameters; conversely, PF correlated with CT parameters, except for overinflated tissue (Table 2). D-Dimer and PF showed an inverse correlation (r −0.46, p-value 0.001). Moreover, NLR, D-Dimer, and LDH correlate mainly with high-density tissues. All three parameters resulted as predictors of non-aerated tissue in linear regression analysis, although NLR only was independently associated with higher non-aerated lung tissue in a multivariate regression analysis (p-value 0.0001).

Predictive Validity of Quantitative CT Parameters and Association with Outcomes
All CT parameters except total lung weight and overinflated tissue showed good predictive validity in a univariate analysis. When CT parameters were compared (avoiding linearly dependent covariates: well-aerated tissue and lung volumes showed high VIF, i.e., more than 10), non-aerated lung tissue only showed independent predictive validity.
NAw% resulted an independent predictor of disease progression compared to NAv% (both with low VIF) and other variables (i.e., D-dimer, NLR, and PF, which resulted predictive of poor prognosis in the univariate analysis) ( Table 3).
When patients were categorized for median NAw%, Kaplan-Meier's analysis showed a significant risk of primary and secondary outcomes in patients with higher NAw% ( Figure 5).

Discussion
Our historical prospective study included hospitalized non-ICU COVID-19 patients. Our analysis highlights some critical findings: (i) Quantitative lung CT allows accurate staging of the severity of COVID-19 pneumonia beyond the extension of infiltration on which current severity score is based; (ii) Risk modeling of non-ICU COVID-19 patients reached the best C-index once it included non-aerated tissue over conventional risk stratification based on clinical, laboratoristic, and respiratory findings.
Different studies have still shown critical advantages of quantitative assessment of COVID-19 [52,53].
First, quantification of lung volumes with lower total and well-aerated lung tissue in patients with disease progression than in patients discharged (as in our case series) turned out to be a key factor in setting some mechanical ventilation parameters [39,49,[54][55][56][57][58][59].
Moreover, quantitative lung CT at admission also resulted effective in the short-term risk stratification of COVID-19 patients for disease progression, i.e., mechanical ventilation and/or death [60,61].
Colombi et al. highlighted the predictive validity of visual semiquantitative or automatic quantitative analysis of the extension of the well-aerated lung on a patient population of 236 [62]. Similarly, Lanza et al. showed the advantages of high-density lung quantification (including non-aerated and poorly aerated areas) in stratifying COVID-19 patients on 222 participants [63].
However, volumetric quantification suffers from the inability to specifically differentiate areas with different gas/tissue ratios, i.e., to adequately differentiate non-aerated (0% gas) and poorly aerated (50% gas and 50% tissue) lung tissue.
This discrimination is logically critical. Progressive respiratory failure develops in many COVID-19 patients, with severe illness reported in 14% of cases, in some case series [1].
Although different clinical risk factors were associated with severe and critical COVID-19, clinical evolution often remains unpredictable, and an accurate biomarker of disease aggressiveness is lacking [64].
Clinical variability depends on different interactive pathological mechanisms.
In fact, COVID-19 showed different patterns, which distinguished the pulmonary pathobiology of COVID-19 from that of equally virus infection [65,66].
In confirmation of this, in a previous work of Chiumello et al., a weak correlation between the venous admixture (cause of hypoxemia) and non-aerated tissue suggested that the major component of the venous admixture in COVID-19 pneumonia is the ventilationperfusion mismatch rather than the true right-to-left shunt due to nonventilated consolidated tissues (unlike what was previously observed in typical ARDS) [39,42,43,49,73,74].
This evidence is consistent with the presence of ground-glass opacities (with or without localized pulmonary consolidations) and the immune-driven intrapulmonary thrombosis since the early phase of the disease in a lung with preserved mechanics [39,[48][49][50]73].
In our case series also, PF and high-density tissue showed only a weak correlation. In contrast, an almost moderate correlation was found between D-Dimer and PF and between D-Dimer and high-density tissues, supporting the hypothesis of the presence of a hypercoagulable state.
The predictive relationship observed between D-Dimer, LDH, and NLR with nonaerated tissue in our results suggests that the latter is more likely to occur with a severe inflammation status in a later phase of the disease when consolidation and fibrotic-like changes prevail [80].
Probably due to these considerations, non-aerated tissue showed an independent predictive validity on patient outcome and was the best index of disease progression, compared to the other tissues and covariates.
This evidence was also confirmed when tissue volume and weight were compared, showing the independent predictive validity of the latter probably related to a greater capability in discriminating disease activity over the extension of disease. Supporting this, the C-index of the model including non-aerated tissue reached 97%, while it did not vary substantially when extension (volume) only was considered.
Therefore, modeling the patient's risk by quantifying non-aerated tissue over clinical and respiratory findings results highly effective.
This marker resulted effective despite the following: (i) The inability of CT to differentiate consolidation and potentially recruitable atelectasis within the non-aerated lung tissue [81]; (ii) The early CT pattern showed in our case series, considering the observed percentage of pulmonary involvement and NAw (%).
(iii) The dichotomy between the dynamic nature of COVID-19 disease and the capability of CT to provide only anatomical information at the time of acquisition. However, the predictive model resulted effective especially when both clinical information and quantitative CT parameters were considered, i.e., in patients with moderate-to-severe symptoms who required hospitalization, thus confirming its prevalent role in the patients' management beyond conventional risk prediction.
This study had several limitations: (i) it was a single-center retrospective analysis; (ii) there was a limited sample for the inter-rater agreement for software-based quantification; however, this is an automatic quantification based on densitometric values and is therefore expected to remain high also with an increased sample; (iii) there was a limited sample size.
An accurate modeling of prediction of disease aggressiveness continues to be a cornerstone in the approach to COVID-19.
Quantitative lung CT provides wide advantages in disease stratification of patients with moderate-to-severe symptoms requiring hospitalization, offering significant insight into COVID-19 lung disease through the automatic detection of high-density areas.
Notably, non-aerated lung tissue mainly showed a strong predictive validity for disease progression to mechanical ventilation or death and therefore should be included in the prognostic model of COVID-19 patients.
Author Contributions: Conception and design: P.P. and M.M.P.; administrative support: A.G., E.D.C. and C.M.; provision of study materials or patients: G.P., A.C. and A.G.; collection and assembly of data: G.P., A.C., A.I., C.A., F.B. and F.S.; data analysis and interpretation: P.P., M.M.P. and F.M.; manuscript writing: all authors. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.

Institutional Review Board Statement:
The study complies with the Declaration of Helsinki principles, and the Institutional Review Board has granted its ethics approval.

Informed Consent Statement:
A reasonable effort was made to obtain informed consent from all patients due to the retrospective nature of the analysis.