Computed Tomography-based Lung Residual Volume and Mortality of Patients With Coronavirus Disease-19 (COVID-19)

Abstract

Rationale and Objectives: 

To assess the effect of computed tomography (CT)-based residual lung volume (RLV) on mortality of patients with coronavirus disease 2019 (COVID-19).

Materials and Methods: 

A single-center, retrospective study of a prospectively maintained database was performed. In total, 138 patients with COVID-19 were enrolled. Baseline chest CT scan was performed in all patients. CT-based automated and semi-automated lung segmentation was performed using the Alma Medical workstation to calculate normal lung volume, lung opacities volume, total lung volume, and RLV. The primary end point of the study was mortality. Univariate and multivariate analyses were performed to determine independent predictors of mortality.

Results: 

Overall, 84 men (61%) and 54 women (39%) with a mean age of 47.3 years (±14.3 y) were included in the study. Overall mortality rate was 21% (29 patients) at a median time of 7 days (interquartile range, 4 to 11 d). Univariate analysis demonstrated that age, hypertension, and diabetes were associated with death (P<0.01). Similarly, patients who died had lower normal lung volume and RLV than patients who survived (P<0.01). Multivariate analysis demonstrated that low RLV was the only independent predictor of death (odds ratio, 1.042; 95% confidence interval, 10.2-10.65). Furthermore, receiver operating characteristic curve analysis demonstrated that a RLV ≤64% significantly increased the risk of death (odds ratio, 4.8; 95% confidence interval, 1.9-11.7).

Conclusion: 

Overall mortality of patients with COVID-19 may reach 21%. Univariate and multivariate analyses demonstrated that reduced RLV was the principal independent predictor of death. Furthermore, RLV ≤64% is associated with a 4-fold increase on the risk of death.

Coronavirus disease 2019 (COVID-19) outbreak was initially reported in Wuhan, China, in December 2019.^1,2 The pathogen was identified as a novel enveloped RNA beta-coronavirus with phylogenetic similarities with the severe acute respiratory syndrome coronavirus (SARS-CoV).^3,4 Therefore, the new discovered virus was named as severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2).⁵

In general, COVID-19 is an acute, mild to moderate disease of the lungs that may progress to a fatal course secondary to a massive alveolar injury and progressive respiratory failure.^2,6 As of June 3, 6,287,771 cases have been confirmed, with 379,941 deaths. Locally, the Americas, including North and South America, reported a total of 2,949,455 cases with 165,311 deaths, representing the leading regional source of new cases diagnosed with COVID-19 worldwide.⁷ Similarly, the mortality in the region has progressively increased with higher numbers concentrated in the United States, Brazil, and Mexico.⁷

Definitive diagnosis of COVID-19 is usually made using a reverse-transcriptase-polymerase-chain-reaction (RT-PCR) assay, with a reported sensitivity ranging between 42% and 83%.^8–12 Recently, to overcome the limitations of molecular tests, chest computed tomography (CT) has been added to the diagnostic workup of patients with COVID-19, especially if RT-PCR assays are not available.^11,13,14 A previous publication by the Radiological Society of North America (RSNA) and the Fleischner Society demonstrated different recommendations about the use of imaging on the assessment of COVID-19 based on different scenarios. According to their recommendations, chest imaging should be used to assess disease worsening and in patients with moderate to severe COVID-19 features regardless of test results. In addition, in a resource-constrained setting, when point-of-care COVID-19 testing is not available, chest imaging is advised to rapidly assess patients with moderate to severe features consistent with COVID-19 and high pretest probability.^15,16

Previous studies have reported that peripheral and basilar predominant distribution of ground-glass opacities with and without consolidations are the most frequent imaging findings of COVID-19.^14,17–21 In addition, chest CT imaging is used to assess severity of the disease and progression. A recent study demonstrated that extensive opacities of the lung parenchyma were more frequently seen in patients with severe disease. In this study, however, parenchymal opacities extension was assessed using subjective visual methods. Often, these evaluations are limited by evaluators’ experience. In addition, the association between disease severity, comorbidities, and chest-CT findings was not assessed.²²

To overcome the limitations of visual scoring systems, CT-based quantitative methods using deep learning algorithms have been proposed to contribute to the diagnosis and severity assessment of COVID-19. Preliminary results have demonstrated that the quantification of lung opacities is associated with disease severity on hospital admission.^23–25 However, the effect of CT quantitative parameters on disease progression and mortality has not been assessed. Similarly, current studies have not analyzed the effect of imaging results on mortality controlling results by clinical factors such as comorbidities and respiratory symptoms. Therefore, the purpose of the study was to assess the effect of quantitative chest CT image parameters based on lung segmentation analysis, defining normal lung volume (NLV), pathologic lung volume, and residual lung volume (RLV), on mortality of patients with COVID-19.

MATERIALS AND METHODS

Study Population

After institutional review board (IRB) approval was obtained, a single-center, retrospective study of a prospectively maintained database was performed using an IRB-approved written informed consent waiver. All patients included in the study were admitted to the hospital following the institutional protocol, as follows: all patients were initially assessed at the COVID-19 screening emergency room. Emergency room assessment was performed using the National Early Warning Score (NEWS) 2²⁶ and Quick SOFA (qSOFA) scales.²⁷ In general, patients with NEWS 2 score higher than 4 and/or qSOFA score higher than 1 were admitted to the hospital. All hospitalized patients underwent RT-PCR testing and noncontrast-enhanced chest CT scanning. Patient characteristics including age, sex, comorbidities, clinical severity scores, and baseline laboratory results including RT-PCR results were collected from electronic medical records.

Patients were excluded from the study if baseline chest CT was performed outside our institution, or baseline CT images had respiratory artifacts that did not allow adequate lung segmentation, or patients had evidence of other viral infection and/or patients without complete clinical information required to assess the primary end point.

CT Scanning Protocol

Noncontrast-enhanced chest CT imaging was performed using 2 CT scanners (Siemens SOMATOM drive and Siemens SOMATOM emotion scanners, Siemens Healthineers, Germany). All patients were scanned in supine position at the end of a full inspiration. The scanning range included from the apex to the base of the lungs. Acquisition and reconstruction parameters were adjusted as follows: 80 to 120 kV tube voltage with tube current of 50 to 350 mAs, pitch of 0.99 ∼1.22 mm, matrix of 512×512, slice thickness of 5 mm, field of view of 350×350 mm, and lung window with a width between −1500 and −700 HU. Imaging reconstructions were performed with a 1-mm slice thickness without interstice gap, using filtered-back-projection reconstruction-based SOMATOM definition. Multiplanar reconstructions were based on axial 1-mm images. Additional prono and expiration images were not obtained.

CT-based Lung Segmentation

Lung segmentation postprocessing analysis was performed using the Alma Medical workstation version 5.0. Segmentation analysis was performed using lung window images with 1-mm slice thickness. Each lung was segmented individually using manual whole-organ volume-of-interest definition. Automated segmentation tool was selected with a reference attenuation range to attempt a thresholding-based segmentation (attenuation thresholds).²⁸ A visual confirmation was performed by an experienced radiologist (3 y of experience assessing chest images) to adjust boundary identification of the lungs. Then, the semi-automated tool was selected to perform a region-based delineation²⁸ of the suborgan pulmonary structures such as airways, blood vessels, lesion boundaries, and/or delineation of lung fields. Once segmentation was approved by the reviewer, a quantitative volumetric analysis was performed defining the following parameters:

Homogenous Nonpathologic Region-NLV

Automated segmentation tool was selected, and a reference attenuation range between −1000 and −600 HU was designated. Vascular structures, airways, and pathologic opacities were excluded using a semi-automated tool. Segmentation boundaries were verified by an experienced radiologist (Fig. 1).

FIGURE 1: A 57-year-old male patient with COVID-19 was evaluated with a nonenhanced chest CT scan at baseline. Multiplanar reconstruction was performed, and lung segmentation of the normal lung parenchyma was performed. Automated and semi-automated segmentation tools were applied, excluding abnormal ground-glass opacities, vascular structures, and airways.

Lung Opacities Segmentation-Lung Opacities Volume (LOV)

Initial automated segmentation was attempted using thresholding-based methods. A reference attenuation range between −500 HU and 20 HU was selected. Then, the semi-automated option was selected to perform a region-based segmentation to adjust lesion boundaries (Fig. 2).

FIGURE 2: A 55-year-old female patient with COVID-19 was evaluated with a nonenhanced chest CT scan at baseline. Multiplanar reconstruction was performed, and lung segmentation of the abnormal lung opacities was performed. Automated and semi-automated segmentation tools were applied to define exclusively abnormal ground-glass opacities. Major vascular structures and airways were excluded.

After completion of a region of interest-based segmentation, a 3D reconstruction was performed to calculate volumes automatically by the workstation. Volumes from each side were added to calculate total NLV and LOV (Fig. 3).

FIGURE 3: 3D reconstruction of a region of interest-based lung segmentation. Lung volumes were calculated automatically by imaging workstation.

Total lung volume was calculated adding NLV+LOV. Then an RLV was calculated as follows:

RLV

Residual volume was expressed as percentage (%) and calculated as follows:$\mathrm{RLV}=\frac{\mathrm{NLV}}{\left(\mathrm{NLV}+\mathrm{LOV}\right)}\times 100.$

Study Outcomes

The primary end point of the study was mortality. Secondary end points included overall survival analysis, need of ICU admission, the use of mechanical ventilation, and length of hospital stay.

Statistical Analysis

Categorical variables were expressed as counts (percentages), and continuous variables as mean±SD or median with interquartile range (IQR). Lung segmentation parameters did not demonstrate normal distribution using the Shapiro-Wilk and Kolmogorov-Smirnov tests. Therefore, logarithmic transformation was attempted. However, normal distribution was not obtained. Univariate analysis, therefore, was performed using nonparametric analysis using the Mann-Whitney U test. Categorical variables were analyzed using the χ²/Fisher test. Variables with significant associations on univariate analysis were included in a multivariate analysis model using the logistic regression analysis. Odds ratios (OR) and 95% confidence intervals (95% CI) were calculated to determine predictors of mortality. Kaplan-Meier curve and log-rank test were used for survival analysis.

The predictive ability of significant factors was investigated using the receiver operating characteristic curve (ROC) method. We computed the area under the curve index (ROC) and evaluated the optimal cutoff point predictive of mortality together with sensitivity and specificity.

All statistical analyses were performed using the SAS 9.4 software (SAS Institute, Cary, NC).

RESULTS

Over a 30-day period, between March 22 and April 22, 2020, a total of 162 patients were reviewed. After inclusion and exclusion criteria assessment was performed, 24 patients (14.8%) were excluded from the study. Finally, 138 patients (84 men [61%] and 54 women [39%]) with a mean age of 47.3 years (±14.3 y) with COVID-19 were included in the study. Overall, 38 patients (28%) had history of hypertension, 33 (25%) obesity, 33 (25%) diabetes, and 27 (20%) had history of smoking. Median length of hospital stay was 8 days (IQR, 5 to 11 d). ICU admission was required for 27 patients (19%). Mechanical ventilation was necessary for 38 patients (28%). Overall mortality rate was 21% (29 patients) at a median time of 7 days (IQR, 4 to 11 d). The demographic and clinical characteristics of the patients are listed in Table 1.

TABLE 1 - Baseline Characteristics of Patients With COVID-19, According to Study Groups

Variables	Overall (N=138)	Patients Who Died (N=29)	Patients Who Survived (N=109)	P*
Age (y)	47.3±14.3	55±11.8	45±14.3	0.0009
Sex (male)	84 (61)	18 (62)	66 (61)	0.9
Comorbidities
Hypertension	38 (28)	14 (48)	24 (23)	0.007
Diabetes	33 (24)	14 (48)	19 (19)	0.001
COPD	5 (4)	2 (7)	3 (3)	0.3
Cardiovascular	9 (7)	2 (7)	7 (7)	0.9
Obesity	33 (24)	8 (28)	25 (23)	0.5
Symptoms
Fever	80 (58)	17 (58)	12 (41)	0.9
Cough	111 (82)	27 (93)	84 (79)	0.08
Fatigue	74 (54)	20 (69)	54 (51)	0.08
Dyspnea	95 (71)	25 (86)	70 (64)	0.04
Diarrhea	28 (20)	2 (7)	26 (24)	0.03

^*P<0.05 is considered to indicate statistical significance.

COPD indicates chronic obstructive pulmonary disease.

After univariate analysis, patients who succumbed were older (55 y [±11.8 y] vs. 45 y [±14.3 y], P<0.01), 14 (48%) had history of diabetes (P<0.01), and 14 (48%) had hypertension (P=0.007). Baseline characteristics by study groups are listed in Table 1.

CT Findings of COVID-19

Baseline chest CT was obtained at a median time of 7 days (IQR, 3.5 to 9 d) from symptom onset. Among patients with COVID-19, ground-glass opacities were observed in 110 patients (79%). In addition, septal thickening was observed in 66 patients (47%), with 53 patients (38%) demonstrating crazy-paving pattern. A total of 102 patients (74%) demonstrated bilateral distribution of chest CT findings.

Lung Segmentation Analysis Results

Lung segmentation analysis demonstrated a median total lung volume of 2215 mL (IQR, 1653 to 2821 mL), with a median LOV of 523.8 mL (IQR, 146 to 918 mL), median NLV of 1432 mL (IQR, 911 to 2432 mL), and a median RLV of 71% (IQR, 58% to 94%).

Lung Volumes and Mortality

Patients who died were found to have a significantly lower NLV compared with patients who survived (median 979 mL [IQR, 772 to 1324] vs. 1722 [IQR, 1070 to 2905], P<0.01). Similarly, the RLV in patients who died was 49.6% (IQR, 37.1% to 67.4%) versus 81.4% (IQR, 60.9% to 97%) (P<0.01) in patients who survived (Fig. 4).

FIGURE 4: Box and whiskers graphs illustrate differences on normal lung volume (mL) (A) and residual lung volume (%) (B) between patients who survived and died. Patients who succumbed demonstrated lower NLV and RLV.

Multivariate Analysis

Age, diabetes, and hypertension were found to be associated with death on univariate analysis. Similarly, NLV and RLV were found to have a significant association with death. Based on general recommendations for multivariate models, multiple logistic regression models were created with up to 3 independent variables to adjust models to the total number of end point events (29 deaths).²⁹ In the first model, diabetes, hypertension, and RLV were included. RLV was the only independent predictor of death, with an OR of 1.042 (95% CI, 1.02-1.065) (P=0.0002) (Fig. 5A and Table 2). Similarly, when age and RLV were analyzed together, only RLV demonstrated to be an independent predictor of death (OR, 1.040; CI, 1.018-1.063; P=0.004) (Fig. 5B and Table 2). Similarly, a logistic regression analysis was performed including RLV and baseline symptoms including dyspnea and fatigue. Only RLV was found to be an independent predictor of death (OR, 1.04; 95% CI, 1.02-1.066) (P<0.01).

FIGURE 5: Multivariate analysis results (logistic regression analysis and ROC curve). A, Logistic regression analysis including RLV, hypertension, and diabetes. B, Logistic regression analysis including RLV and age. C, Logistic regression analysis including RLV and baseline symptoms (fatigue and dyspnea). D, Diagnostic performance of RLV at multivariate analysis and receiver operating curve method.

TABLE 2 - Logistic Regression

Predictor	OR (95% CI)	P
Model 1
RLV (%)	1.042 (1.02-1.065)	0.0002
Hypertension	0.39 (0.13-1.15)	0.08
Diabetes	0.53 (0.17-1.6)	0.26
Model 2
RLV (%)	1.04 (1.018-1.063)	0.0004
Age (y)	0.96 (0.933-1.005)	0.08
Model 3
RLV (%)	1.043 (1.02-1.066)	0.0002
Fatigue	0.74 (0.27-2.04)	0.57
Dyspnea	0.79 (0.21-2.92)	0.73

Independent predictors of death.

To assess the diagnostic performance and to determine cutoff points of RLV on mortality, a receiver operating method was performed. The area under the curve values and optimal cutoff points are reported in Table 3. The effect of dichotomized calculated cutoff point of RLV on mortality was assessed on a logistic regression model. An RLV lower than 64% (adjusted OR, 4.8; 95% CI, 1.9-11.7) (P<0.001) was selected based on ROC results (Table 3).

TABLE 3 - Diagnostic Performance of RLV Independently Associated With Death and Accuracy of Calculated Cutoff Points

Predictor	AUC	Optimal Cutoff Point	Sensitivity, %	Specificity, %	Adjusted OR (95% CI)	P
RLV (%)	0.79	64%	85.3	50	4.83 (1.99-11.72)	0.0005

AUC indicates area under the curve.

Survival Analysis

Overall survival rate at 5, 10, and 15 days was 93%, 75%, and 61%, respectively. The survival rate in patients with RLV equal or lower than 64% was 85%, 61%, and 47% at 5, 10, and 15 days, respectively, compared with patients with higher RLV, who demonstrated a survival rate of 98%, 88%, and 74% at 5, 10, and 15 days, respectively (P=0.01) (Fig. 6).

FIGURE 6: Survival analysis. A, Overall survival at 15 days. B, Kaplar-Meier curve of patients by RLV ≤64% and >64%.

DISCUSSION

Fatal course of hospitalized patients with COVID-19 can be as high as 21%. As reported in previous studies, several factors are associated with death, including comorbidities. According to our results, baseline NLV and RLV are important morphologic parameters associated with death at a median time of 7 days. In fact, multivariate analysis and survival analysis demonstrated that reduced residual volume was the most significant predictor of mortality. Furthermore, a RLV equal or lower than 64% resulted in a 4-fold increase in the risk of death, with a significant survival reduction at 15 days.

Worldwide, mortality of COVID-19 has continuously increased since the beginning of the outbreak. According to the 135th report by WHO, overall mortality reached 6%. Currently, the Americas represent the leading region of new COVID-19 cases worldwide, with the highest number of fatalities concentrated in the United States, Brazil, and Mexico.⁷ To our knowledge, this is the first study that reports in-hospital mortality in Latin America.

In-hospital mortality rates secondary to COVID-19 may range between 1.4% and 50% in different regions of the world depending on illness severity.^1,30,31 According to previous reports, deceased patients had a median age of 60 years, frequently with history of diabetes and/or hypertension (±30%).^1,30–32 In our study, in-hospital mortality rate was 21%. Patients who died were 10 years older than patients who survived (55 vs. 45 y). However, in our cohort, patients were younger than previous reports. Likewise, deceased patients had history of diabetes and hypertension (48%). In our study, however, when comorbidities, baseline symptoms, age, and lung segmentation results were included in a multivariate model, only CT-based lung volumes were found to be independent predictors of death.

In previous studies, lung CT findings have been studied to assess severity of COVID-19. Qualitative and quantitative parameters have been proposed. In a previous study, Yu et al²² reported that patients with severe disease had more extensive opacities of the lung parenchyma with interlobular septal thickening, air bronchograms, and pleural effusions. This study, however, was limited by the use of subjective evaluations, being influenced by evaluator’s experience.

To overcome this limitation, quantitative parameters have been proposed using deep learning algorithms. Lu Huang et al²⁵ used a deep learning–based instrument to quantify lung opacities. Segmentation results were reported as percentages. According to their results, opacity percentages were higher among patients with severe disease. Similarly, Cheng et al²⁴ developed a multi-scale convolutional neural network to assess lung segments with pathologic opacities. Quantitative and semi-quantitative methods were used to assess disease extension. Quantitative results were reported as percentages. According to their results, extension percentages were correlated with disease severity indexes and abnormal laboratory results.^24,25 One of the limitation of these studies is that the association among imaging results; clinical factors such as comorbidities, age, and/or symptoms; and adverse outcomes was not analyzed. Furthermore, the number of studies assessing the effect of baseline lung CT results on mortality is limited.

In our study, normal lung parenchyma and opacities extension were quantified and reported as volumes. To assess disease extension, residual NLV was calculated and expressed as percentage. According to our results, reduced normal lung parenchyma volume and low RLV were associated with death. Furthermore, in a multivariate model, we found that low RLV was the only independent predictor of death. Other clinical parameters such as comorbidities, symptoms, and age failed to demonstrate significant association with mortality on multivariate analysis.

In our analysis, a RLV cutoff point of 64% was determined with a sensitivity of 85.3% to predict death. Patients with RLV ≤64% were found to have a 4-fold increase on the risk of death when compared with patients with higher RLVs. Likewise, 15-day survival rate was significantly lower among patients with low RLV compared with patients with RLV higher than 64% (47% vs. 74%, respectively).

There are several limitations of the study. First, the retrospective nature of the study restricts the inclusion of patients with COVID-19 who required hospital admission. To avoid bias, the collection of data and recruitment of patients was prospective following a standardized protocol. All patients were admitted to the hospital following NEWS 2 and qSOFA results. Both scales are approved and widely used by large medical centers to screen patients with possible COVID-19 and to assess the need of admission. Second, although commercially available software was used for lung segmentation, the current version still requires radiologists’ verification. In addition, imaging segmentation software is not fully available for all healthcare providers or radiology departments. Third, the lung segmentation assumes that the volume of lung opacities is a surrogate for COVID-19 involvement; however, no histologic confirmation was available to confirm CT findings. Fourth, the association with other clinical parameters, including blood oxygenation laboratories and inflammatory response parameters, was not assessed in this study. We expect to publish a supplementary study that expands current results including inflammatory response parameters and blood gas. Therefore, further studies are needed to replicate and validate our results, specifically the effect of RLV cutoff points assessing the probability of death.

In conclusion, in-hospital mortality of patients with COVID-19 may reach 21%. Univariate and multivariate analyses demonstrated that reduced NLV objectively calculated by lung segmentation analysis is an independent predictor of death. Furthermore, RLV equal or lower than 64% is associated with a 4-fold increase in the risk of death, with a significant decrease on the survival rate at 15 days when compared with patients with higher RLV.