Positional Therapy and Regional Pulmonary Ventilation: High-resolution Alignment of Prone and Supine Computed Tomography Images in a Large Animal Model

All studies were approved by the Animal Care and Use Committee of the University of Pennsylvania, Philadelphia, Pennsylvania. Detailed methods are provided in Supplemental Digital Content 1 (https://links.lww.com/ALN/C459). The study is a reanalysis of images included in a previously published article,¹⁶  with uniquely different image analysis methods, nonoverlapping results, and new clinical implications.

Five Yorkshire pigs (32.4 ± 2.7 kg) were anesthetized with intramuscular injections of ketamine and xylazine. Pigs were orally intubated, and anesthesia was maintained by continuous intravenous infusion of ketamine and midazolam through a peripheral catheter. Supplemental doses were given to assure immobility during computed tomography scanning. Peripheral oxygen saturation, heart rate, and body temperature were monitored continuously. The femoral artery was catheterized for blood pressure monitoring and arterial blood gas measurement.

Animals were ventilated with a custom-built mechanical ventilator. Tidal volume 8 ml/kg, respiratory rate 15 breaths per minute, and inspired fraction of oxygen 0.6 were constant throughout the experiment. Pigs then received 2 ml/kg hydrochloric acid (pH 1.25) injected in the endotracheal tube over two equally divided doses with the animals alternately in the right and left lateral position. Pigs were then stabilized supine for 60 min with PEEP 10 cm H₂O. Imaging was performed before injury (healthy baseline) as well as after injury and stabilization. Blood gases were obtained after each image acquisition. Airway pressure was monitored using a fiberoptic sensor placed proximally to the endotracheal tube. Tidal volume was recorded from the ventilator display.

Computed tomography scans were acquired with a Siemens (Germany) SOMATOM Force scanner. The settings were as follows: 120 kVp, 200 mAs, pitch 0.95, slice thickness 0.75 mm, collimation 57.6 × 0.6 mm, and estimated dosage 3 to 5 mSv. All images were reconstructed to a resolution of 1 × 1 × 1 mm. End-inspiratory images were obtained during 5-s-long inspiratory pauses, applied during ventilation in both supine and prone positions with PEEP 5 and 10 cm H₂O. Each combination of PEEP and position was applied in random order and maintained for 10 min before imaging, which was shown in pilot studies to assure sufficient stabilization of aeration redistribution. In the prone position, the pigs lay unsupported on their abdomen with their forelegs flexed and their hindlegs extended at the hips. Animals did not receive neuromuscular blocking agents, but sedation was adjusted to achieve immobility, which was confirmed on computed tomography scans by the absence of motion artifacts near the diaphragm.

Our computed tomography analysis (fig. 1) measured changes in computed tomography density (i.e., aeration) and dimensions (i.e., spatial deformation) in individual small units (voxels) of tissue. For this purpose, images were registered, i.e., reshaped (warped) and aligned to each other to match the position of each voxel on shared spatial coordinates. This registration corrected for the effects of body position on lung and chest wall geometry. The outlines of the lungs were segmented (separated) from nonpulmonary tissue.

We registered corresponding supine and prone images (fig. 1) using Advanced Normalization Tools.^12,17  The entire registration code is shown as text in Supplemental Digital Content 1 (https://links.lww.com/ALN/C459). During Step 1 of the registration, the outline (mask) of the lungs was segmented from nonpulmonary tissue using previously published methods.¹⁸  In Step 2 of the registration (fig. 1), the prone lung outline was aligned to the supine outline, resulting in a “warped-prone” outline with identical outer geometry to the supine lungs. During Step 3 of the registration, all the pulmonary structures contained within the outlines of the prone and supine computed tomography scans were aligned, yielding a warped image of the prone lungs for each study condition.¹¹  The quality of the registration was assessed in three ways. First, overlap between the warped-prone and supine lung outlines was quantified using the Dice similarity coefficient. The Dice coefficient was 0.988 ± 0.002 before injury and 0.983 ± 0.006 after injury, suggesting optimal overlap of the lung outlines between warped-prone and supine images. Second, trained operators measured minimal differences in the spatial coordinates of anatomical landmarks between warped-prone and supine images, suggesting conserved spatial orientation of parenchymal structures (see figure, Supplemental Digital Content 2, https://links.lww.com/ALN/C460, showing the results of this analysis). Third, we compared regional density distributions between original and warped prone images (see table, Supplemental Digital Content 3, https://links.lww.com/ALN/C461, showing that registration preserved lung density distributions).

In Step 4, we measured small-scale aeration changes by calculating the differences in x-ray film attenuation, measured in Hounsfield Units, between tissue voxels of warped-prone and supine images. Each voxel’s Hounsfield Units value quantifies the increase or decrease of gas content in that same tissue unit. We quantified body position–related changes in the dimensions of each tissue voxel as its corresponding Jacobian.¹⁹  A Jacobian is a unitless metric of tissue deformation calculated during image registration. When comparing corresponding warped-prone and supine images, a voxel with a negative determinant of the log Jacobian represents contraction, whereas a positive value indicates expansion incurred by the same lung tissue unit. Detailed analysis of lung recruitment was performed by identifying and mapping the voxels that shifted from the high density, nonaerated range (–100 to + 100 Hounsfield Units) to lower density ranges²⁰  when comparing the warped-prone to the supine images. Derecruitment was measured as voxels that increased density to higher than –101 Hounsfield Units when turning prone. Whole-lung contents of gas and tissue were estimated by density measurements on the nonregistered images.²¹ 

K-means clustering was performed on the voxel distribution of Hounsfield Units (x-axis) and of Jacobian (y-axis) to characterize three different groups (reinflation, limited-change, and deflation). The number of clusters (k = 3) was chosen during preliminary analysis because it maintained physiologically plausible inflation responses and stable distributions among animals. Both Hounsfield Units and Jacobian were normalized by standard score [(x – µ)/σ] to prevent bias due to the value dispersion between x- and y-axis (μ is the mean of the population and σ is the standard deviation of the population). For each cluster, the percentage of total lung mass and volume it occupied and its mean values of Hounsfield Units, Jacobian, and computed tomography intensity were calculated in each condition. Gravitational and nongravitational patterns of cluster distribution were separately investigated by partitioning the lung images into 10 horizontal and 10 vertical bins of equal mass.

No power analysis was performed due to the retrospective design of the study. Descriptive statistics were performed on the clusters identified by K-means, characterizing their size as a percent fraction of total lung volume and mass, and their values of mean computed tomography intensity, Hounsfield Units, and Jacobian. All continuous data were reported as mean ± SD. Repeated-measures ANOVA was performed to compare the cluster distributions across spatial compartments. Two-tailed paired t tests were performed for relevant cluster distribution comparisons, and to compare physiologic values between data obtained at healthy baseline versus injury, prone versus supine, and lower versus higher PEEP. An experiment-wide P < 0.05 was selected as the threshold for statistical significance. Data was reformatted and prepared in Jupyter Notebook using Python 3.6. Statistical analysis was performed using RStudio 1.2.5001 (R Foundation for Statistical Computing, Austria).

The imaging results are a reanalysis of a previously published computed tomography dataset from a pig model of mild lung injury. All animals survived and were included in the analysis. Hydrochloric acid injection caused mild lung injury, with a decrease in Pao₂ from 251.5 ± 50.4 to 161.1 ± 116.9 mmHg (P = 0.047), and with a decrease in respiratory system compliance from 20.1 ± 3.1 to 14.6 ± 4.2 ml/cm H₂O (P = 0.046) at PEEP 5 cm H₂O in the supine position (see table, Supplemental Digital Content 4, https://links.lww.com/ALN/C462, showing blood gases and respiratory mechanics data). After injury, the prone position increased Pao₂ at both PEEP levels, but it did not affect compliance. Higher PEEP did not modify gas exchange or compliance.

Representative baseline computed tomography images obtained before acid-induced lung injury are shown at inspiration with ventilation at PEEP 5 cm H₂O (fig. 2A) and at PEEP 10 cm H₂O (fig. 2B). The supine, prone, and corresponding warped-prone images are shown in the left panels of figure 2. The corresponding subtraction maps (fig. 2, middle panels) show negative Hounsfield Units values (blue) in the dorsal lung regions, indicating increased gas content from supine to prone. In contrast, density increased (positive Hounsfield Units values, in red) in the ventral lung, indicating locally reduced aeration when prone. The Jacobians extracted during registration of prone to supine images were mapped and displayed next to the subtraction maps (fig. 2, middle panels). Negative Jacobian values (blue) were visible in ventral lung regions, indicating tissue contraction in the prone position in the regions where Hounsfield Units were positive. In the dorsal regions, positive Jacobian values (red) dominated, indicating expansion in the regions where density decreased.

Postinjury images at PEEP 5 and 10 cm H₂O (fig. 3, A and B, left panels) show that, after registration, supine and warped-prone images differed only in the distribution of pulmonary densities, which were more prominent in the supine position. Similar to healthy baseline, the regional Hounsfield Units values at both PEEP levels were generally negative (improved aeration) in the dorsal lung and positive (decreased aeration) in the ventral regions (fig. 3, A and B, middle panels). The Jacobian maps also displayed similar behavior (ventral contraction, dorsal expansion) to that in healthy lungs.

Cumulative frequency distributions (including all animals) of Hounsfield Units and Jacobian values were more heterogeneous after injury (fig. 3, A and B, right panels) than in healthy conditions (fig. 2, A and B, right panels). Mean Hounsfield Units values were near zero in all the tested conditions, suggesting that the prone position had minimal net effect on whole-lung gas content. This was confirmed by the stability of mean lung density and whole-lung gas volumes between prone and supine positions (see table, Supplemental Digital Content 5, https://links.lww.com/ALN/C463, showing the results of computed tomography density analysis in the nonprocessed images).

Cluster analysis assigned lung tissue voxels to clusters with distinct characteristics (fig. 4). The following cluster characterization represents general trends, derived from analysis of all conditions (table 1; see also figure, Supplemental Digital Content 6, https://links.lww.com/ALN/C464, with descriptive statistics of computed tomography–derived characteristics of each cluster). The most ventral cluster (blue, fig. 4) occupied on average (including all conditions studied) 39.0% of the total lung volume and 29.3% of total lung mass, and was characterized by positive Hounsfield Units and negative Jacobian values, indicating loss of aeration and tissue compression (deflation). The midlung cluster (black, fig. 4) occupied a larger fraction of lung tissue (47.0% of total lung volume and 47.6% of total mass, on average for all conditions), with small average values of Hounsfield Units and Jacobian suggesting limited response to position. The third cluster in the dorsal lung (red, fig. 4) had negative Hounsfield Units and positive Jacobian values, indicating reinflation, and was the smallest cluster (all-condition averages of 14.0% of total lung volume and 23.1% of total lung mass). Cluster percent distributions, Hounsfield Units, and Jacobian values were not significantly affected by injury versus healthy status. After injury, the reinflation cluster was smaller at PEEP 10 cm H₂O than at PEEP 5 cm H₂O (P = 0.039). Regional quantification of cluster distribution in 10 vertical and 10 horizontal bins of equal mass showed both ventro-dorsal and craniocaudal gradients (fig. 5). The 10 horizontal bins (fig. 5A) were partitioned evenly into dorsal and ventral compartments, while the vertical bins (fig. 5B) that interfaced the diaphragm were grouped into a caudal versus a cranial compartment. Cluster distribution significantly differed between both ventro-dorsal (F[1, 4] = 103.9, P < 0.001), and cranio-caudal compartments (F[1, 4] = 117.5, P < 0.001). Of the reinflation cluster, 86.4 ± 8.5% was located within the dorsal compartment versus 13.6 ± 8.5% in the ventral compartment (P < 0.001). Of the reinflation cluster, 63.4 ± 11.2% was contained in the caudal, peridiaphragmatic bins versus 36.6 ± 11.2% (P < 0.001) in the remaining bins.

Figure 6 displays representative postinjury maps of recruitment and derecruitment due to prone position at end-inspiration at both PEEP 5 cm H₂O and PEEP 10 cm H₂O, with group mean ± SD values shown below each map. Prone positioning recruited 64.5 ± 36.7 g of tissue (11.4 ± 6.8% of total lung mass) at low PEEP and 49.9 ± 12.9 g (8.9 ± 2.8% of total mass) at higher PEEP. Of recruited voxels, 59.0 ± 13.0% were in the caudal region. Derecruitment was relatively small in both conditions. We also registered PEEP 5 and 10 cm H₂O images to each other for each position (left two columns) in order to visualize tissue with restored aeration at higher PEEP. The images show that this and position-related recruitment were quantitatively similar. Both recruitment and derecruitment were low in healthy lungs (see table, Supplemental Digital Content 7, https://links.lww.com/ALN/C465, showing recruitment and derecruitment data in all conditions).

We directly visualized the effects of positioning on each unit of lung tissue in paired prone–supine computed tomography scans of ventilated pigs. We then identified which regions of the lung responded to the prone position with reinflation, deflation, or limited change. These patterns followed both gravitational and nongravitational distributions: reinflation was mostly concentrated in the caudal lung region near the dorsal portion of the diaphragm (figs. 4 and 5). Recruitment of nonaerated tissue contributed to reinflation in this same region after injury (fig. 6).

Using image registration, we were able to match lung structures in paired images to provide a detailed quantification of position-related changes in regional lung volume and gas content (figs. 2 and 3). After this step, cluster analysis grouped tissue units according to changes in density (aeration) and dimensions (expansion or contraction). We then mapped the spatial distribution of the three clusters, which had distinct responses to prone positioning: ventral deflation (increased density and contraction), limited tissue response in the midlung, and dorsal reinflation (decreased density and tissue expansion). We used K-means clustering,²²  with its simplicity and low computational complexity, to generate nonsubjective regions of interest that shared similar patterns of the two markers we presented in this work (density change and tissue deformation). This method is suitable for classifying an image into different regions when the number of clusters is known for particular human anatomy and/or physiology.^23,24  We postulated a three-cluster model, as it would reflect expected simple airspace responses. We are confident in this model, as it resulted in stable classifications between animals and conditions. Furthermore, the opposite density changes measured in the reinflation and deflation clusters were physiologically plausible.

Our methods allow us to quantify the effects of compressive phenomena on the distribution of aeration in the prone versus supine position. The observed ventral contraction and dorsal expansion of lung tissue corroborates studies proposing that attenuation of the gravitational gradient of lung density in prone ARDS patients³  is due to redistribution of the forces that deform the lung,^7,25,26  causing more uniform pleural and transpulmonary pressures than in the supine position.^4,27  Additionally, we measured predominant reinflation and recruitment (after injury) in the dorso-caudal lungs, indicating localized positional responses near the diaphragm. The concurrent craniocaudal and ventro-dorsal aeration patterns confirm the coexistence of gravitational and nongravitational forces. This finding is in accordance with the results of a detailed ex vivo analysis of lung tissue, which showed that gravity does not dominate the topographic effects of position on lung function.⁸  In fact, position-related changes to perfusion and ventilation were prominent in the dorso-caudal lung region of ventilated pigs,⁸  which could be explained by focal compression and deflation in the same regions when animals are in the supine position. Such regional patterns result from the reconfiguration of the diaphragm profile that occurs with position due to changes in abdominal pressure.⁴  These effects of positioning on the diaphragm are known,^28,29  but here we proved a first clear, direct visualization of their implications on regional lung inflation.

Our approach has additional advantages over the analysis used in previous imaging studies.³  Registration of paired images yields detailed maps of density and tissue deformation, which are otherwise impossible to obtain. Because our voxel by voxel analysis does not impose constraints on the analysis (gravity or arbitrary density thresholds), the distribution of the clusters shows complexity that would be otherwise underappreciated. In fact, we detected voxels of different clusters intermingling with each other in the coronal maps (fig. 4), representing gravity-independent heterogeneity in response to position.

Prone position had negligible effects on the average density and gas volume of the whole lungs, results that are in discordance with studies where prone positioning improved lung capacity.³⁰  However, other studies reported stability of lung volumes when comparing prone to supine positions in animals³¹  as well as in humans.³²  Similar to human studies,^33,34  we did not observe improvements in respiratory system compliance in the prone position. This absence of effect was likely related to compression of the ventral chest wall reducing chest wall compliance, which offset improvements in lung compliance.^35,36  We did not support shoulders and hips in the prone position, which restrains lung expansion and limits ventral inflation,^2,32  and likely caused density to increase in a large region of ventral lung (fig. 4).

In the setting of unchanged gas content, gas exchange improvements in the prone position were likely related to focally improved recruitment and ventilation-perfusion matching in the dorso-caudal region of the lung. Studies by others have quantified lung recruitment by prone position,^32,37  but registering paired images allowed us to map the reopening of collapsed tissue with voxel-level detail. As shown in figure 6, tissue voxels recruited by the prone position were dorsally distributed and were predominantly located within the dorso-caudal region occupied by the reinflation cluster. Dorsal recruitment was not matched by an equivalent amount of ventral derecruitment in this model. The net balance in favor of recruitment was compatible with the mild injury and dorsal distribution of the lesions, with less tendency for lung collapse in the ventral lung regions. Furthermore, density was more homogeneous between clusters in the prone compared to supine position (see panels C and D in Supplemental Digital Content 6, https://links.lww.com/ALN/C464, showing cluster densities in each position). Similar to previous studies,³⁸  such even aeration indicated reduced tendency for ventilation loss in the prone position. Oxygenation and recruitment improved in the prone position at both levels of PEEP; however, improvement in aeration due to higher PEEP (fig. 6) was not associated with improved oxygenation. This discrepancy was probably related to more favorable distributions of perfusion and of ventilation observed in the prone position versus higher PEEP.³⁹  However, a larger gradient between tested PEEP level could have revealed whether spatial patterns of reinflation are superimposable when comparing higher PEEP versus prone positioning.

Our study has several limitations. First, the use of an animal model limits the extrapolation of our results to human ARDS. While the vertical distribution of computed tomography density in pigs is qualitatively similar to that reported in humans,³²  lung and thoracic geometries may be different. We studied a mild model of lung injury, but we showed significant radiologic changes and sizable oxygenation and recruitment responses to prone positioning. Furthermore, previous animal studies showed that prone positioning delays the progression of mild injury,^5,16  supporting the relevance of our results. The small number of animals also limits statistical power, but we were nevertheless able to effectively describe established gravitational and novel nongravitational effects of prone position on lung tissue. Registration is a key component of our analysis, and its use to study body position is novel. Validation of this methodology is ongoing within the field,⁴⁰  but we showed that peripheral pulmonary outlines, major airway structures, and density distributions were accurately superimposed when comparing warped-prone to supine images, indicating fidelity in the conservation of spatial coordinates between positions. We have displayed similar validations of these methods in our previous studies of tidal inflation in registered inspiratory–expiratory images.⁴¹  Our study lacked esophageal manometry, and we cannot discriminate the effects of positioning on pulmonary and thoracic mechanics. Last, while we allowed 10 min of stabilization before imaging after each change in PEEP and body position, longer periods of stabilization may have elucidated greater changes in gas exchange and respiratory mechanics.

Our study is unique in that we were able to directly visualize improvements in small-scale aeration and match them with tissue expansion and local recruitment due to prone position. Although we did not directly test their biologic implications, these patterns are important because the effects of positional therapy on patient mortality are not explained by better gas exchange.⁴²  Therefore, imaging characteristics may be implicated in predicting injury progression and response to therapy. The current study shows that the beneficial effects of prone position on lung recruitment and aeration are concentrated in the same dorso-caudal region. It is therefore likely that the presence of opacifications in the dorsal lung and their responses to prone position (recruitment, reinflation) are key indicators of the therapeutic and protective effects of this therapy. Noninvasive imaging tools such as lung ultrasound⁴³  can assess localized responses in the dorso-caudal lung and possibly predict outcome and personalize care in individual patients.

We directly measured and mapped the effects of prone position on lung density and small-scale inflation, and found that improvements in both metrics were concentrated in the dorso-caudal region of the lung. These localized changes may help interpret positional therapy outcomes.

The authors thank the Center for Advanced Computed Tomography Imaging Services (CACTIS) at University of Pennsylvania (Philadelphia, Pennsylvania), especially Ms. Sheila Murdaugh (American Registry of Radiologic Technologists), who spent many weekends with us imaging the subjects.

This work was supported by National Institues of Health (Bethesda, Maryland) grants R01-HL139066, R01-HL124986, and R01-HL137389.

Dr. Cereda received grant support from Hillrom (Indiana). Dr. Gee received grant support from University of Electronic Science and Technology of China, International Neuroinformatics Coordinating Facility, Asian Society of Magnetic Resonance in Medicine, ShanghaiTech University, China. The remaining authors declare no competing interests.