A quantitative analysis of extension and distribution of lung injury in COVID-19: a prospective study based on chest computed tomography

For each chest CT scan, twenty images, evenly spaced along the cranial-caudal axis between the apex and the diaphragmatic dome, were selected for analysis [21, 22]. In order to describe the progress of lung lesions along the subpleural-to-core direction inside lung parenchyma, three concentric subpleural regions of interest were defined: from pleural surface to 1 cm depth; from 1 to 2 cm depth; and from 2 to 3 cm depth. To distinguish between hypodensities related to lung injury and hypodensities due to dependent atelectasis and pleural effusion, each lung was divided into four quadrants. Consequently, three functional regions were identified: non-dependent (external non-dependent + internal non-dependent), dependent (external dependent + internal dependent) and external (external non-dependent + external dependent). Concentric ROIs and lung regions were therefore combined for regional analysis (Fig. 1).

Each voxel composing the CT images was characterized by a CT number expressed as Hounsfield units (HU). Following a previous convention [23], four lung compartments were defined: hyperinflated (HU between − 1000 and − 800), normoinflated (HU between − 800 and − 500), hypoinflated (HU between − 500 and − 100) and non-inflated or atelectasis (HU between − 100 and + 100). The extension of each lung compartment was expressed as percentage of the total lung volume analyzed in a given chest CT slice (defined in Figs. 2 and 3 as % of total lung volume). The extension of pathologically altered lung was defined as the sum of hypoinflated and non-inflated (atelectatic) parenchyma [24, 25] distributed in not-dependent or external regions. For each of the 20 slices selected per CT scan, lung volume, gas volume and lung weight have been calculated according to established methods [21, 22, 26] applying the following equations:

When the voxel volume is expressed as cubic centimeters [cm³], the volume of gas is in milliliters [ml] and the tissue weight is in grams [g]; voxel attenuation is expressed as Hounsfield units [HU].

Applying the above-mentioned equations, total lung volume, total gas volume and total lung weight have been then calculated from the sequence of CT scans, following the interpolation method reported by Ball et al. [22] and by Reske et al. [21]:where N is the number of slices, t is the slice thickness, f the distance between slices, and M_i the lung mass in the ith slice. This equation yields the total mass of the lung Mlung. This general interpolation method is also able to compute total lung volume and total gas volume, when the mass Mi is appropriately substituted by, respectively, the lung volume or the gas volume in the ith slice.

The mentioned equation was used for computing lung volume, gas volume and lung weight for the entire lung as well as for each lung aeration compartment. The latter have been reported both as absolute values and as percentage of the total lung parenchyma.

Patients’ subgroups were defined based on dichotomous variables: 1) exposure to mechanical ventilation before CT scan (yes/not); 2) exposure to a mean tidal volume higher/lower than 6 mL/kg predicted body weight (PBW); 3) duration of spontaneous breathing for more/less than 14 days from the onset of symptoms; 4) plasma ferritin lower/higher than 1000 µg/L (normal plasma ferritin range: 12 to 150 µg/L). Spontaneous breathing, before invasive ventilation, was defined as all types of assisted and not-assisted noninvasive ventilation. The selected thresholds for spontaneous breathing duration and for plasma ferritin were based on previous literature [27].

Continuous variables were reported as medians and interquartile ranges (IQR). Friedman's test was used to detect statistical differences in HU distribution among the three subpleural layers and the whole lung parenchyma. Pairwise comparisons were made when analysis of variance detected a significant difference inside the tested groups. The adjustment for multiple comparisons was applied by using the Bonferroni method. A Wilcoxon rank sum nonparametric test was used to evaluate statistical differences between subgroups defined by dichotomous variables. Spearman correlation test was used to analyze correlations between lung compartments and clinical data. Pairwise linear correlation coefficients (R_s) and p-values were collected and interpreted based on α = 0.05.

More detailed Materials and methods are reported as Additional file 1.

The distribution of lung compartments among the analyzed lung regions is reported in Fig. 2 and Additional file 2: Figures S1–S2. Hyperinflated parenchyma was more represented in non-dependent regions and equally distributed at different distances from the pleura. The dependent lung regions were characterized by more extended hypodensities (hypo- and non-inflated lung) compared to non-dependent lung regions. This at the expenses of normoinflated lung regions (Additional file 2: Figure S2 and Tables S1–S2). Normoinflated, hypoinflated and non-inflated compartments were more common in lung regions close to the pleura (0 to 1 cm below pleura) compared to more central ones (2 to 3 cm). This was true independently from the selected quadrant and was found in not dependent as well as in dependent lung regions, hence, showing that the distribution of hypo- and non-inflated compartments was not only influenced by gravitational forces (Fig. 2, Additional file 2: Figures S1–S2 and Tables S1–S2).

The statistically significant correlations are presented in Additional file 2: Tables S4 and S5. A direct correlation (positive R_s) was found between the amount of non-inflated lung and respiratory rate at ICU admission. An inverse correlation (negative R_s) was found between the amount of hypoinflated lung and pH at ICU admission. The hypoinflated lung correlated with days of symptoms, duration of mechanical ventilation, renal replacement therapy (RRT), prone positioning and muscle relaxation. Measured tidal volume (mL/kg PBW) was correlated with the extension of hypoinflated lung in patients on mechanical ventilation. A positive cumulative fluid balance and plasma D-dimer were correlated with the extent of injured lung. Moreover, injured lung was inversely correlated with peripheral oxygen saturation, and normally ventilated lung was inversely correlated with partial pressure of carbon dioxide.

COVID-19-associated lung injury showed a specific sub-pleural distribution (Fig. 2 and Additional file 2: Figure S1) coexisting with a varying involvement of the core areas of the lung. This confirmed what has been previously described in a qualitative manner [7, 8, 20]. However, our results did not support a strict division between two respiratory phenotypes [7], but rather a more gradual evolution from a subpleural to a more central distribution of lung injury [8]. The prominent subpleural distribution was confirmed for both dependent and non-dependent lung regions and so unrelated to a gravitational distribution, which means we could exclude confounding by atelectasis and pleural effusion that typically affect the dependent regions of the lung. As already known, normoinflation was mainly distributed in subpleural regions [28]. To date, the reasons for a primary distribution of COVID-19-associated lung injury on subpleural regions are not fully clarified. Gattinoni and coworkers [7] speculated about the possibility that subpleural regions are areas of interface between lung structures with different elastic proprieties, hence converging higher stress and strain [29]. A complementary explanation is provided by the physics of the deposition of inhaled viral particles. Particles of the size of the virion (0.15 μm), deposit mainly in the terminal alveolar region [30], being scarcely affected by inertial impaction or gravitational sedimentation on airways walls. The following phase is characterized by an inflammatory pattern that spreads centripetally, possibly following the direction imposed by the ciliar beats on mucus and the watery periciliary liquid [31]. The higher proportion of injured central regions of the lung after 14 days of spontaneous breathing has been demonstrated in our study (Fig. 3).

A direct correlation between the extension of injured areas and the exposure to mechanical ventilation or to a non-protective tidal volume (Additional file 2: Tables S4 and S5) further corroborated the analysis based on dichotomous variables (Fig. 3). Our data suggested that protective ventilation with limited tidal volumes is of particular importance in COVID-19-associated acute respiratory failure. More liberal mechanical ventilator settings have been proposed for COVID-19-associated acute respiratory failure [9] although not supported by univocal evidence and in contrast with the necessity of keeping a tidal volume equal or lower than 6 mL/kg PBW [32, 38]. Our results do not support this more liberal ventilatory approach. Recent studies indicate the importance of a protective ventilation in ARDS patients [14, 39], including COVID-19 patients [14]. Even a short exposure to high intensity mechanical ventilation can potentially be harmful and associated with increased mortality, independently from the severity of the initial lung injury [14, 39].

We found a direct correlation between injured lung and patients’ pronation or muscle relaxation (Additional file 2: Table S4). This can be easily explained by the fact that more severe forms of SARS-CoV2 related acute respiratory failure are more likely in need of pronation [40] and more often undergo muscle relaxation as standard of care [41]. Nine (39%) out of 23 patients included in the study underwent pronation with improved oxygenation during their intensive care stay.

The direct correlation found between the onset of lung injury and a cumulative fluid balance (and need of RRT) confirmed the importance of avoiding fluid overload in patients with COVID-19-associated acute respiratory failure. The need for RRT is an indicator for acute kidney injury and most likely for a positive fluid balance. Previous investigations have shown that a positive fluid balance is associated with the increased risk of death, more days of mechanical ventilation and a longer ICU-stay in patients with ARDS [42, 43].

We did not find any correlation between the amount of lung injury and the tested inflammatory biomarkers (list in Additional file 2: Table S4). This supported previous results [44] suggesting that the severity of COVID-19-associated acute respiratory failure is poorly related to commonly used inflammatory biomarkers. Nevertheless, as stated before, the subgroup of patients showing ferritin levels higher than 1000 µg/L the day of the CT scan was associated with a significantly larger lung injury (Fig. 3).

Although not able to demonstrate a direct link between plasma D-dimers and thrombotic burden, we found that high plasma D-dimers levels correlated with lung injury detected by CT. A recent study by Grasselli and coworkers [14] shows that high plasma D-dimers are correlated to high 28-day mortality in COVID-19-associated acute respiratory failure. A number of changes toward a pro-thrombotic state, including an increased level of D-dimers and activation of the lectin pathway of complement activation, have been observed in patients with severe COVID-19 [45, 46]. The correlation found between lung injury and D-dimers was consistent with a spread phenomenon of pulmonary micro-thrombosis observed in histological studies [47]. Nonetheless, the activation of a pro-thrombotic state at the alveolar level is one of the events characterizing all forms of ARDS and leading to the intra-alveolar fibrin deposition, vascular injury and micro-thrombi occurrence [48].

The present study has different limitations. The subgroup of patients selected for the radiological investigations was sicker compared to the whole cohort, showing higher 20 day mortality (36% vs 16%) and SAPS3 (57 vs 51) (Table 1). Positive end-expiratory pressure levels and other ventilatory settings selected during the acquisition of chest CT scans varied depending on best clinical practice, adding an element of heterogeneity in the comparisons of different CT scans (Table 2). The CT scans were performed on clinical indication. The consequent non-standardized approach derived from the good clinical practice guidelines in force during the first pandemic wave when local and international routines for management and transportation of critically ill COVID-19 patients were still under development. Patients numerosity as well as the number of chest CT scans usable for the current analysis was limited by the necessity of including only chest CT scans acquired without contrast agent. The contrast mean can variably alter the HU scale depending on its blood concentration and the regional distribution of tissue. The inclusion of images containing contrast would have introduced a further source of heterogeneity whose resolution would have required rely upon assumptions and analysis whose complexity is beyond the scopes of the present contribution. The lack of a control group with classical ARDS made the comparison between COVID-19-associated acute respiratory failure and other forms of ARDS impossible. The correlations tested in this study did not necessarily imply causality. The choice of 14 days for estimating the influence of timing of mechanical ventilation is based on the evidence that in different studies this is the time point within which the most of the patients are intubated [27, 49].