Human bodies
This prospective study was approved by the Institutional Review Board and was conducted between November 2015 and July 2018. Human bodies were transferred to the Institute of Forensic Medicine at coroner’s inquest. Due to the mode of inclusion, no preselection of the human bodies according to certain criteria, e.g., the presence of specific lung diseases, was possible. Externally visible conditions causing a significant impairment of the normal thoracic anatomy and signs of advanced decomposition were exclusion criteria. Imaging was performed before autopsy no longer than 36 h after death with bodies cooled to slow decomposition. The imaging was not part of the forensic analysis. Altogether, nine bodies (3 females, age range 52–88 years; 6 males, age range 60–83 years) were imaged. Airway pressure was kept constant (20–25 mbar) during x-ray dark-field imaging by endotracheal intubation and mechanical ventilation.
X-ray dark-field imaging
The setup was previously described [18, 21]. The employed three-grating arrangement is asymmetric [periodicity of G0, G1, and G2 is 68.72 μm, 8.73 μm, and 10 μm, respectively; inter-grating distances: d(G0–G1) = 1.60 m, d(G1–G2) = 0.25 m]. Gold heights for all gratings range between 150 and 200 μm. As already described [23], the shadow of G1 is directly projected onto G2. G1 and G2 are tiled to each cover an area of 40 × 2.5 cm2. The tiling procedure has been already described [24]. All gratings are mounted on a swing pivoting around the focal spot. Acquisition is performed via fringe-scanning, yielding a field of view of 32 × 35 cm2. The source (MRC 200 0310 ROT-GS 1004, Philips Medical Systems, Hamburg, Germany) is an actively cooled tungsten rotating anode and was operated at 70 kVp, where a mean visibility of 31% was achieved. A flat panel detector (Pixium RF 4343, Trixell, Moirans, France) was used. Source and detector remain stationary during acquisition. Imaging was performed in supine position with anterior-posterior beam setup. Acquisition time was 40 s. For additional information, see Additional file 1: Figure S1.
CT imaging
Human bodies were imaged in supine position on a 256-slice CT unit (Brilliance iCT, Philips, Amsterdam, Netherlands). High-resolution chest CT was performed in craniocaudal direction with 128 × 0.625 mm collimation and 0.383 pitch. Tube voltage was 120 kVp. Mean tube current was 537 mA. CT images were reconstructed with iDose4, a hybrid iterative reconstruction technique (Philips, Amsterdam, Netherlands), at level 2 in axial, coronal, and sagittal view with a slice thickness of 3 mm, 1024 × 1024 matrix, and 350-mm field of view.
Data acquisition and processing of x-ray dark-field imaging
The fringe-scanning method [25] was used for data acquisition: a fringe pattern is induced on the detector by detuning inter-grating distances. Acquisition while moving the pattern across the sample produces images of the same features at multiple relative grating shifts.
Signal extraction was performed using the least-squares minimisation of an image formation model similar to the one presented in [26]. To correct for visibility reduction due to beam-hardening, a correction algorithm comparable to the method presented in [27] was used.
The dark-field radiographs were low-pass-filtered (2D Gaussian filter kernel, σ = 3.2 pixels). This reduces noise levels by a factor of ~ 11.3 for white noise, leading to a visual impression more similar to conventional radiography. Although low-pass filtering obscures small features, these are nearly undetectable in the unfiltered images due to the high noise levels. The used kernel size was found to be an acceptable trade-off between image impression and resolution. No filtering was applied to conventional radiographs.
Reader study
Visual image analysis was independently performed by three residents with 3 (F. M.), 5 (A. S.), and 5 (D. D.) years of experience in chest imaging on a clinical Picture Archiving and Communication System workstation. For training purposes, the dark-field radiograph of human body 4 was presented before the reading session to demonstrate low and high dark-field signal intensity. In the first reading session, window settings were fixed to allow optimal comparison of low and high dark-field signal intensities in dark-field radiographs and opacification in conventional chest x-rays to avoid influence of individual windowing. A linear mapping between grey values and logarithmic visibility reduction ratios, − ln(V/V0), was used. Window level and width were set to 8,500 and 5,000, respectively. Converting these numbers back to physical quantities, this means that a logarithmic visibility reduction ratio of -0.268 corresponds to “black,” and a value of 0.343 corresponds to “white.”
The nine x-ray dark-field radiographs had to be graded separately one after the other without the possibility to compare or change gradings. Next, the conventional x-rays were presented. On each image, the left and right lung were divided into three regions of equal height, upper, middle, and lower zones, using the apex and the costodiaphragmatic recess as anatomical landmarks. Dark-field signal intensity and degree of transmission (or opacification) of the upper, middle, and lower zones of the left and right lung were graded on a 6-point (0–5) ordinal scale (Fig. 1). For the dark-field signal intensity grading, “0” represents no (dark area in the radiograph) and “5” a high (bright area in the radiograph) dark-field signal. “1–4” represent intermediate dark-field signal intensities (Fig. 4 for comparison). For the transmission grading, “0” represents no transmission or hyperattenuation like in the clinical case of a pleural effusion where no ventilated lung parenchyma is visible. “5” represents a normal, healthy lung with high transmission or hypoattenuation. “1–4” represents intermediate transmission grades (Fig. 7 for comparison).
The reading session was repeated after 4 weeks.
In a separate reading session, the readers independently graded image quality for right and left lung on a 6-point ordinate scale: 1 = not diagnostic, 2 = sufficient, 3 = satisfactory, 4 = good, 5 = very good, and 6 = excellent. As standardised image quality criteria for dark-field radiographs do not exist, the readers were instructed to evaluate the following aspects: symmetrical reproduction of the thorax, reproduction of the whole lung, and presence of artefacts interfering with the grading of pulmonary dark-field signal intensity (e.g., vertical streaking artefacts, dark-field signal from bony structures). For transmission images, the “European guidelines on quality criteria for diagnostic radiographic images” [28] were applied wherever possible considering imaging of a human body in supine position. In this setting, readers were free to change window/level values to optimise individual image impression.
Correlation of dark-field and transmission radiography with CT findings
As there exists no data on x-ray dark-field imaging features of human lung pathologies, we performed a CT scan of each human body to correlate findings in chest CT images with signal changes in dark-field and transmission radiographs. CT images were reviewed by an attending radiologist with 10 years of experience in chest radiology (A. A. F.) using axial, sagittal, and coronal reconstructions. Pulmonary findings and extrapulmonary findings with a potential effect on dark-field signal intensity were recorded. Apart from septal thickening, the extent of pulmonary findings was visually quantified for every lobe in 10% intervals. For pleural effusions, the maximum width in anterior-posterior direction was measured in centimeters. Other findings were qualitatively recorded. CT findings were correlated with the visual assessment of dark-field signal strength in a descriptive model.
Statistical analysis
Statistical analysis was performed using GraphPad Prism 7 for Mac OS X (Version 7.0d, GraphPad Software Inc., USA) and R version 3.4.4 (R Foundation for Statistical Computing, Vienna, Austria). For ordinal data, median and range are presented and dot plots with medians and 95% confidence interval (based on the Hodges-Lehmann method) are shown. Intraobserver and interobserver agreement of dark-field signal and transmission grading were evaluated using weighted Cohen κ with squared weights. Cohen κ coefficients with values < 0 were regarded as poor, 0–0.20 as slight, 0.21–0.40 as fair, 0.41–0.60 as moderate, 0.61–0.80 as substantial, and 0.81–1.00 as (almost) perfect agreement, according to Landis et al. [29]. Differences in distributions of dark-field signal and transmission grading for the upper, middle, and lower zones were tested separately for the right and left lung and each reader using the Friedman test. If Friedman test indicated a significant (p < 0.050) association between region and dark-field signal, Wilcoxon matched-pairs signed-rank test was performed for pairwise comparisons of regions of each lung for each reader. The correlation of dark-field signal with transmission grading was tested with Spearman’s rank correlation coefficient for each region of lungs and each reader. Differences in the grading of image quality between the left and right lung for dark-field and transmission radiographs, respectively, and between dark-field and transmission radiographs for left and right lung, respectively, were tested with Wilcoxon matched-pairs signed-rank test.