Detection of inspiratory recruitment of atelectasis by automated lung sound analysis as compared to four-dimensional computed tomography in a porcine lung injury model

Background Cyclic recruitment and de-recruitment of atelectasis (c-R/D) is a contributor to ventilator-induced lung injury (VILI). Bedside detection of this dynamic process could improve ventilator management. This study investigated the potential of automated lung sound analysis to detect c-R/D as compared to four-dimensional computed tomography (4DCT). Methods In ten piglets (25 ± 2 kg), acoustic measurements from 34 thoracic piezoelectric sensors (Meditron ASA, Norway) were performed, time synchronized to 4DCT scans, at positive end-expiratory pressures of 0, 5, 10, and 15 cmH2O during mechanical ventilation, before and after induction of c-R/D by surfactant washout. 4DCT was post-processed for within-breath variation in atelectatic volume (Δ atelectasis) as a measure of c-R/D. Sound waveforms were evaluated for: 1) dynamic crackle energy (dCE): filtered crackle sounds (600–700 Hz); 2) fast Fourier transform area (FFT area): spectral content above 500 Hz in frequency and above −70 dB in amplitude in proportion to the total amount of sound above −70 dB amplitude; and 3) dynamic spectral coherence (dSC): variation in acoustical homogeneity over time. Parameters were analyzed for global, nondependent, central, and dependent lung areas. Results In healthy lungs, negligible values of Δ atelectasis, dCE, and FFT area occurred. In lavage lung injury, the novel dCE parameter showed the best correlation to Δ atelectasis in dependent lung areas (R2 = 0.88) where c-R/D took place. dCE was superior to FFT area analysis for each lung region examined. The analysis of dSC could predict the lung regions where c-R/D originated. Conclusions c-R/D is associated with the occurrence of fine crackle sounds as demonstrated by dCE analysis. Standardized computer-assisted analysis of dCE and dSC seems to be a promising method for depicting c-R/D. Electronic supplementary material The online version of this article (10.1186/s13054-018-1964-6) contains supplementary material, which is available to authorized users.

2 analyses). Body temperature was measured via rectal probe and maintained between 37.5 and 38.5 °C by body surface warming. After completing the protocol, the animals were euthanized in deep anaesthesia by injection of 40 mmol potassium and 20 mg/kg propofol.

Details of Study Protocol
After BLH measurements, model lung injury was induced -according to, yet slightly modified from, a previous report on repetitive lung lavages [1] -to study c-R/D. respectively. If at ZEEP the PaO 2 /FIO 2 ratio was still > 300 and at a PEEP of 15 cm H 2 O the PaO 2 /FIO 2 ratio was > 450, another lavage was performed. This procedure was repeated until the PaO 2 /FIO 2 had decreased to < 300 at ZEEP, or if maximal recruitment had become compromised (PaO 2 /FIO 2 < 450 at PEEP 15 cm H 2 O). Subsequently, the animals were stabilized for another 30 minutes and analogous to BLH; measurements during model lung injury were recorded at randomized PEEP levels (LAV 0; LAV 5; LAV 10; LAV 15). The animals remained in dorsal recumbency for the entire experiment.
The duration of the protocol was as follows: approximately 180 minutes for induction of anaesthesia and instrumentation; 80 minutes for BLH measurements (four PEEP levels at approximately 20 minutes each); subsequent induction of lung injury, ranging from 70 to 180 min, and 80 min for measurements after lavage. 3

Time-synchronization of all recording systems
To time-synchronize the different measuring devices, a local area network between the different recording devices was set up, and using a time-server unit, we adjusted all system times to synchronize with the time stamp of the CT-scanner. Moreover, we were able to verify this approach, because the time stamp set by the time-server (when CT acquisition started) correlated perfectly with the change in frequency characteristics above 2000 Hz induced by CT acquisition noise (for more details see paragraph "Influence of CT acquisition noise on lung sound recordings"). Although all attempts were made to synchronize the different recording methods, we cannot preclude a minimal temporal offset caused by CT image processing time. However, this seems of minor relevance, as a normal CT scanner makes images available at the same magnitude as the acquisition time.

Preliminary experimental tests
Before carrying out the study, the influence of the surrounding noise on the attached piezoelectric contact sensors and the influence of these upon the radiologic imaging quality were addressed. Using a noise-absorbing mat wrapped around the subjects, pre-testing showed that the recorded raw data sound waveforms were not noticeably affected.

Influence of acoustic sensors on CT imaging
To identify the impact of the acoustic sensors (Meditron, ASA, Norway) on computed tomography (CT) imaging quality, multiple test scans were performed using a plastinated lung (i.e., a plastic casting of a real lung). The most affected lung cross-section with the sensors attached was identified and compared to the identical lung cross-section after removal of the sensors. We found a bias of up to 40 Hounsfield units (HU) for mean lung densities (MLDs) in CT imaging. Fig. S1 shows exemplarily the impact of the acoustic sensors attached. Moreover, we performed a post-hoc analysis of the 4DCT data sets from two piglets to assess the impact the metallic sensors might have had in regard to the different lung regions analyzed, and in regard to lung movement during respiration. Here we found that the impact of the metallic sensors did not significantly differ between the different lung regions of interest, when analyzing the entire lung stack with a longitudinal coverage of 8 cm (nondependent, central, dependent lung). Further, when looking at the dependent lung (where cyclical recruitment and derecruitment took place), no significant differences were found in the noise caused by the acoustic sensors in the end-expiration lung stack versus the noise in the end-inspiration lung stack. Of course, we observed movement (and changes in noise) caused by mechanical ventilation, but there was nearly no movement in the spinal region of the chest, due to anatomical reasons. This fact, and the fact that we used a saline lavage model injury that produced atelectasis mainly in the dependent part of the lung, might best explain our findings. However, we cannot exclude the possibility that the noise present in the CT signal (which may have been due to these sensors) and the interference caused by movement during mechanical ventilation might have biased the weights of the results and statistics when analysing the complete 4DCT data sets.

Influence of CT acquisition noise on lung sound recordings
We also tested for alterations in the spectra of the recorded lung sounds which might have been induced by CT acquisition noise. Our results showed that CT acquisition noise caused a distinct sound signature above 2000 Hz in frequency, although a noise-absorbing mat was used. However, we want to point out that this exerted no significant influence upon our postprocessed sound parameters (FFT area and dCE), as exemplarily highlighted in Fig. S2. shows the spectrum of the measured lung sounds without the CT running; the orange curve shows the spectrum with the CT running. As one can see, there is an almost perfect agreement between the two spectra -although these were two consecutive breaths.

Details of 4DCT lung segmentation
Lung segmentation and 4DCT analysis was carried out using a specialized software system for analyzing lung CT images, called YACTA, which was developed by one of our authors (OW). The software contains a specific module for semi-automatic and/or full automatic segmentation and quantification of porcine lung 4DCT data.
In our study, one single 4DCT covers 40 CT volumes, each containing 40 CT slicesresulting in 1600 images per 4DCT data set. Due to this large amount of data, a softwarebased support system was mandatory for the evaluation of data.

6
In the following, the 3D segmentation and quantification scheme for a 4DCT data set -as implemented in YACTA -is described in a nutshell. The segmentation of the atelectatic areas of the lung is a very demanding task, since the density values of these areas are only very slightly different from the surrounding structures -sometimes there is no difference in the density values at all. Our study possesses two major advantages: Firstly, we have the CT volumes over the entire breathing cycle (40 volumes). Secondly, the volumes are at the same spatial position. There is of course movement caused by mechanical ventilation, but there is nearly no movement in the spinal region of the chest; the movement increases as one progresses to the ventral regions of the chest.
To take advantage of the first point, we searched for the CT volume where the contrast between lung tissue and surrounding tissue was maximized. To accomplish this, the complete 4DCT data set was loaded into the YACTA software, and the CT volume with the minimal mean density (Minimum Density Stack, MDS) was identified (see Fig. S3A).
Because of this selection, the MDS should be the easiest volume to be segmented out of all volumes in the time series (e.g., please compare Fig. S3A and S3B). In this volume, the following operations were performed for lung segmentation: First, a Gauss filter was applied; then, high-density structures (≥200 HU; most likely representing bone) and extracorporeal areas (≤ -900 HU) were identified and discarded (Fig. S3B). Thereafter, the N6-connected low-density area (≤ -400 HU) covering the highest volume inside the body was identified and labeled as "Lung". This lung segmentation was refined by morphologic hole-filling and closing algorithms. Final lung segmentation results are shown in Fig. S3C. The fully automatic segmentation results of the lung were then manually refined by our radiologists (YY, TA) using a graphic interface and a painting pencil, in case the region did not completely match the anatomical lung region. However, manual changes were only necessary in a few cases.
Thereafter, the following steps utilized the second advantage of our study. All volumes of one 4DCT series were at the same spatial position, and additionally, the atelectatic changes in the pulmonary tissue occurred mainly in the dorsal lung areas (see Fig. S3E) -in which almost no movement takes place throughout respiration. The lung mask generated for the MDS was transferred and adjusted to all other CT volumes belonging to the 4DCT data set.
To adjust the lung borders to correlate most realistically to a specific CT volume, voxels that were within 10 mm distance to the outer border of the previously labeled "Lung" were marked and classified as uncertain voxels (see Fig. S3D). Fig. S3E shows the lung contour (red line) determined in the MDS that was transferred to another CT volume of the 4DCT data set. One can see in Fig. S3E rather nicely that there is almost no movement in the dorsal area, where atelectatic changes are clearly visible. On the other hand, the transferred lung contour does not match very well in the ventral regions of the lung -but there we can still see a higher 7 contrast between lung tissue and surrounding tissue -and thus, automatic adjustment of the lung contour could be easily performed in these more ventral regions.
The actual adjustment of the transferred lung contour was achieved by the following operations: The bounding box for the entire label "Lung" was calculated and the uncertain "Lung" voxels were classified as follows: In the lower 20% of the box, they were simply marked as "Lung" (as in these dorsal regions, nearly no movement in body contours occurred over time). In the middle region of the bounding box, uncertain voxels were marked as "Lung" if HU values were ≤ -200; in the upper region of the bounding box, they were marked as "Lung" if HU values were ≤ -100). Fig. S3F shows the final adjusted lung contour. Further, we performed a regional sub-analysis by dividing the marked "Lung" into equal thirds: The upper third covered the non-dependent lungs areas, the middle third covered the central lung areas, and the lower third the dependent lung areas.
Finally, for each lung region of interest, mean HU and the time course in HU were generated for the mean lung stack and the respective volumes (cm 3 ) of either atelectatic (-300 to 0 HU), poorly ventilated (-600 to -301 HU), normally ventilated (-900 to -601 HU) or hyperinflated (-1024 to -901 HU) lung areas. Fig. S4 shows an example of 4DCT post-processing. An example of the post-processed "FFT area" plot, showing the power spectral density over frequency, is presented in Fig. S5.

Details of spectral coherence method
Our intent in performing this evaluation was to identify the sensors overlying the lung regions where cR/D takes place. The spectral coherence method was chosen based on the global consideration that fine crackle sounds arise through the sudden recruitment of airways and alveoli, resulting in avalanches of adventitious sounds.
To detect the accompanying changes of spectral characteristics in frequencies over time, assessment of the coherence of neighbouring sensors is a commonly used method to identify how well one input signal corresponds to another input signal.
To gain a parameter representing global similarity between the two signals, we summarized the coherence values over the frequency band. Of two neighbouring sensors Lx.My and Lx.M(y+1) the magnitude squared of the mean cross-spectral density Lx.C xy y(f) was assessed. For each frequency, this estimated cross-spectral density with values between 0 and 1 indicates how well one input signal corresponds to another input signal at each frequency -whereby 1 indicates a strictly linear transfer function between both input signals, and 0 indicates linearly unrelated or uncorrelated signals.
For a global interpretation of the dependency over the entire frequency band, the sum ∑(Lx.C xy y(f)) of the Lx.C cxy y(f) function was processed. Thus, for each sensor pair, the result of cross-coherence over the entire frequency band could be reduced to one decimal number/value -with a high value indicating high similarity. Technically, the range for this

Details of 4DCT results
The results for the different PEEP levels of 0, 5, 10 and 15 cm H 2 O in the lavaged lungs (LAV 0, LAV 5, LAV 10, and LAV 15) are shown in Fig. S7. While the graphs on the left represent the overall lung volumes in regard to the lung region, the graphs on the right show the variation in volumes (Δ volume) within the respiratory cycle.   Table S1 and Table S2 shows detailed results of the lung volumes classified as hyperinflated, normally aerated, poorly aerated and atelectatic lung areas, as mean and SD for all animals after lavage injury. Table S1 represents the average lung volumes; Table S2 represents the within-breath changes during ongoing mechanical ventilation.

Details of "dynamic spectral coherence" results
Estimates with standard error (SE) of the pair-wise comparisons of the fitted ANOVA model are presented in Table S3.