X-ray Dark-field Radiography - In-Vivo Diagnosis of Lung Cancer in Mice

Accounting for about 1.5 million deaths annually, lung cancer is the prevailing cause of cancer deaths worldwide, mostly associated with long-term smoking effects. Numerous small-animal studies are performed currently in order to better understand the pathogenesis of the disease and to develop treatment strategies. Within this letter, we propose to exploit X-ray dark-field imaging as a novel diagnostic tool for the detection of lung cancer on projection radiographs. Here, we demonstrate in living mice bearing lung tumors, that X-ray dark-field radiography provides significantly improved lung tumor detection rates without increasing the number of false-positives, especially in the case of small and superimposed nodules, when compared to conventional absorption-based imaging. While this method still needs to be adapted to larger mammals and finally humans, the technique presented here can already serve as a valuable tool in evaluating novel lung cancer therapies, tested in mice and other small animal models.

With approximately 1.5 million deaths a year worldwide, lung cancer is the leading cause of cancer-induced deaths 1,2 . Lung cancer is mostly diagnosed at an advanced stage with poor prognosis and has an average five-year survival rate of 15% 3 . The initial assessment of the lung in clinical routine is often done by conventional chest radiography, however, the sensitivity of detecting small nodules is particularly low. A recent study, for instance, demonstrated that the detection sensitivity drops below 50% for tumors with a mean size of 19 mm 4 . Further, the average size of missed lung nodules on a X-ray radiography is no less than 15 mm 5 . These results may explain why conventional x-ray chest radiography is not optimal as a mortality reducing screening program for (early) detection of lung cancer 6 .
In order to better understand the pathogenesis and develop treatment strategies numerous small-animal studies are currently being performed. In these studies, it is essential to be able to precisely detect and track the development of lung nodules in vivo. Thus, it is possible to assess the impact of therapy over time. Currently, micro-CT and histology are used as the tools of choice for the lung nodule tracking. While histopathology offers an excellent opportunity, it is not suitable for longitudinal studies. Micro-CT can help localize the tumors, but it has a disadvantage of delivering quite a substantial dose to the animal. Accumulation of the dose can therefore bias or even falsify the results of the study 7 .
To address the aforementioned shortcomings, we propose to utilize the recently developed technique called X-ray dark-field radiography for the improved in-vivo assessment of lung cancer 8, 9 . X-ray dark-field is a signal that quantifies the small-angle scattering that occurs in the lung at the air:tissue interfaces. Previous simulation 10 and experimental results [11][12][13][14] have demonstrated that changes to the size and number of the alveoli can be precisely detected with this imaging technique. Thus, recent studies with in vivo mice demonstrated promising potential in detection of pulmonary disorders such as emphysema 15,16 , and pulmonary fibrosis 17 , amongst others 18, 19 . Thereby, the study by Hellbach et al. 16 clearly revealed that the dark-field signal of the lung scales with the size of the alveoli, where in the publication by Yaroshenko et al. 17 it was demonstrated that the signal also decreases with the decreasing number of alveoli. Here we show, using an in vivo small-animal lung cancer model (Fig. 1a) that x-ray dark-field contrast is also highly suitable for the detection of small tumor nodules which could be overlooked at the corresponding conventional image, by monitoring the cancer-caused reduction of the air:tissue volume ratio (Fig. 1b and c) 20 .
Finally, our study is highly significant for both clinical imaging and treatment development, providing a diagnostic and dose-saving framework for the assessment and evaluation of new lung cancer therapeutics and their application in mouse models.

Results
The detection of lung cancer using X-ray dark-field radiography was assessed in two steps. At first a qualitative radiographic investigation, using a transgenic animal model with spontaneous K-ras oncogene-driven growth of lung tumors which closely resembles human lung tumor development, was carried out. Subsequently, a quantitative reader study was conducted, with experienced radiologists to compare detection rates of lung tumor nodules with dark-field versus conventional radiographs.
In-vivo dark-field radiography. Eight K-ras LA2 mutant mice and one wildtype control animal were imaged with a dose of 1.5 mGy using the small-animal scanner (acquisition parameters in Table 1) 21 . Corresponding absorption and dark-field radiograms of three typical animals with lung tumors and one control mouse are presented in Fig. 2 Table 1. Small-animal CT scanner acquisition parameters used for the in-vivo X-ray dark-field radiography lung cancer study.
overlaying the lung, i.e. ribcage, heart shadow and the diaphragm, it is difficult to localize and differentiate these lung lesions on the transmission radiograms. Furthermore, small tumor lesions impose only minor variations to the overall optical density with respect to the surrounding healthy tissue, which renders an unambiguous detection of small nodules extremely challenging. In comparison, the impact of overlaying structures in dark-field imaging is much less severe, since signal arising from the lung -more precisely the alveoli -is dominant in the thorax region. This advantage enhances the assessment of lung lesions, since distortions in the homogeneous (strongly scattering) lung pattern are more easily perceived by the radiologist. Compared to the conventional transmission signal, significantly more lung tumors (especially small nodules) are detectable and visible as (poorly scattering) black voids in the dark-field projections (as indicated by red arrows). In addition, the size and location of the tumors can be more clearly identified and were found to be in agreement with the histopathological work-up.
Reader study: in-vivo dark-field radiography. In order to quantify the increase in diagnostic performance gained through X-ray dark-field imaging, a reader study was conducted. Three experienced radiologists were blinded and both transmission and dark-field radiograms presented separately, with the option of scaling data-sets freely (for more details see Materials & Methods). To achieve a statistical relevance a total of 16 murine lungs (8 cancer-bearing; 8 healthy -7 of these from a previous study 17 ) were considered. The number of detected tumors (true-positives) on transmission and dark-field images for the three readers are presented in Fig. 3, and demonstrate that dark-field imaging significantly outperforms transmission imaging with respect to the assessment of lung nodules, holding true for all three readers and each of the examined mice. We verified the here presented numbers by histology as method of reference. Note that we attribute the distinct inter-reader variance to the fact that the radiologists are trained in the evaluation of human radiographies and were not particularly trained for small-animal investigations. Compared to the transmission radiographs, we found an increase in tumor detection rates, ranging from at least 92% up to a maximum of 267%, when exploiting the dark-field channel. As a matter of fact, in the case of "Mouse 1" (see also Fig. 2) the lung was even rated as healthy on the transmission radiogram by all three radiologists, while 6, 6 and 9 small lesions were detected within the dark-field channel, respectively. Please note that we do not provide a percentage with respect to the overall number of tumors present within the lung, since an exact determination from histological work-up is ambiguous, especially in the case of very small tumors. Finally, our analysis further indicated that dark-field based diagnostics Figure 2. Qualitative benchmark of in-vivo transmission versus dark-field radiography for the assessment of lung nodules. Typical transmission and dark-field radiograms of a control mouse and three mice with lung tumors, alongside with the corresponding histological sections in H&E staining. Small nodules are much easier to detect within the dark-field image (red arrows) than in the attenuation channel, due to a minor impact of overlaying structures and the overall dominant scatter signal of the lung within the thorax. The scale bars correspond to 5 mm and 2.5 mm on the radiograms and histological sections, respectively. The color-scale corresponds to the linear signal of both dark-field and transmission images.
increase the diagnostic specificity of lung tumor assessment, by reducing the number of false-positives within the control group of healthy mice from an overall of 6 to 4 incidences (see also Fig. 3

-bottom).
Exceeding these insights, the reader study further indicated that scatter-based imaging enables a better determination not only of tumor numbers but also shape, size and delineation. This observation is illustrated in Fig. 4, where one of the readers digitally evaluated/sketched the extensions for each of the detected lung nodules. Photographs of the freshly excised lungs were utilized for reference. In case of "Mouse 2" for instance, two separated lesions were originally assumed, based on the transmission signal, in the bottom of the left lung. Evaluating the dark-field signal it is apparent that only one large tumor is present. In case of "Mouse 3", the nodule in the right was estimated as round by the radiologist, while an oval tumor shape was found in the dark-field channel, in agreement with the photograph and pathology.
Nevertheless, Fig. 4 also gives evidence that dark-field radiography is not yet capable of resolving all tumors, as indicated by black circles in the photograph. This is due to the planar nature of a radiographic imaging mode, where very small tumors can potentially be superimposed by other large tumors or the prominent signal of healthy tissue. The largest missed tumor was estimated from histology and measured with a diameter of 0.7 mm.
Hence, in order to evaluate the potential of X-ray dark-field imaging in the assessment of superimposed or very small tumors, we are currently extending our study to tomographic investigations of living mice.

Discussion
Within this study, we presented the very first dose-compatible in-vivo dark-field measurements of mice with lung cancer and found a superior diagnostic performance compared to conventional radiography. Thereby, we verified two main advantages associated with the dark-field signal: first, structures overlying the lung are weakly scattering and thereby do not influence the visibility of the lung. Second, micro-structural differences between healthy and cancerous lung tissue provide intrinsically enhanced scatter contrast.
While these results are very promising, we have to consider that within this proof-of-principle (reader) study none of the lately developed ribcage suppression algorithms had been applied [22][23][24][25] . Although these are yet not applied in daily clinical routine, they yield the potential to increase the quality of lung nodule visualization in transmission radiography as such 22,25,26 . In addition, in translating our findings to clinical imaging, it is uncertain to which degree scatter caused by the human ribcage and spine will corrupt the visibility of the human lung. Therefore, future studies will have to consider the aforementioned algorithms for transmission radiographies to provide a fair comparison. However, a very initial benchmark (Fig. 5) following the method described by von Berg et al. suggests that the inherent micro-structural sensitivity of dark-field imaging strongly outweighs the diagnostic benefits arising from ribcage suppression/correction algorithms only 25 : while one small nodule in the left upper lung (green box) does indeed become detectable in the transmission image after removing the ribcage, another tumor in the upper right lung (blue box) remains hidden in the transmission signal throughout rib-removal, and is reliably detected in the darkfield channel only. Reader study: transmission versus dark-field radiography in-vivo. Based on the transmission and dark-field radiograms, three experienced and blinded radiologists were asked to assess the number of radiopaque tumors in 16 mice (8 tumor-bearing, 8 healthy). A significant improvement in detection rate (truepositives) was found, when exploiting scatter-sensitive dark-field imaging, ranging from an average increase by 92% up to a maximum increase of 267%. Based on the control group our analysis provided evidence, that darkfield radiography additionally enhances diagnostic specificity, by reducing the number of false-positives from an overall of 6 to 4 incidences. . Improved assessment of tumor size, shape and delineation using X-ray dark-field radiography. Transmission and dark-field radiograms of two mice with lung cancer are taken from Fig. 2 and evaluated by a blinded reader. The green ellipses indicate tumors, which were correctly assessed in both imaging channels. The red ellipses indicate tumors which were detected in the transmission image, but misjudged with respect to size or shape, e.g. being mistakenly classified as separated (Mouse 2) or round (Mouse 3). Blue ellipses account for tumors which were correctly assessed within the dark-field radiogram only. Overlooked/missed tumors are indicated on the photograph by black ellipses. Further, we hypothesize that the distinct sensitivity towards pathological changes in the pulmonary micromorphology may be of special interest for the detection of non-solid growing tumors, which are typically radiolucent in conventional radiography in early stages. Here, the tumor slowly proliferates along the alveolar walls reducing the amount of incorporated air and therewith scattering surface, while not significantly changing the optical density of related tissue 20 .
In general, X-ray dark-field imaging requires at least a doubling of dose, due to the presence of an absorber grating. We would like to point out that if a transmission image was acquired with the dose required for x-ray dark-field radiography, its SNR would have been better. However, a factor of two in dose results only in a slight increase of the SNR by a factor of = .
(2) 1 4. Finally, in conventional radiography systems there is an anti-scatter grid behind the sample (which also reduces the flux by approx. 20%), further reducing the difference in SNR between the attenuation images taken with a GI and the images taken with a conventional system. Therefore, we believe the presence of the GI in the imaging system does not influence the quality of the transmission images significantly. Similar imaging results have been previously reported 27 .
The mice were breathing freely during the image acquisition. As no gating was applied, this introduced some blurring to the images. However, given the typical motion amplitude of the murine lung in the range 28 of 200 μm and the resolution of the system 29 of approximately 70 μm, the motion artifacts have been found to be negligible in the past 30 . Therefore, the deviation of the actual tumor size from the estimated one is considered to be negligible.
Our results indicate that x-ray dark-field radiography is a promising candidate for both lung cancer screening and monitoring high-risk human patients, offering a significantly increased cancer detection rate (true-positives), without increasing the rate of false-positives. A fusion of the simultaneously acquired transmission and dark-field signal would further provide the radiologist with anatomical and scatter-sensitive (functional) imaging at a single glance. However, before this technology can be adapted to human imaging, several technical challenges have to be addressed. Experiments presented here were carried out at 35 kVp, whereas for human imaging, energies between 80-120 kVp are typically used. In general, feasibility of x-ray dark-field imaging has been demonstrated for such high energies [31][32][33] . These studies have also demonstrated that while Compton Scattering increases for higher x-ray energies, the G2 grating acts as an anti-scatter grating and allows thus for signal extraction. However, in general the visibility in a grating interferometer is typically reduced by a broad x-ray spectrum. Therefore, future studies should consider how the x-ray spectrum should be shaped to obtain the optimal quality for human imaging. Furthermore, the field of view has to be strongly increased. At the moment, it is limited by the size of the available gratings. Recently, several studies have reported the possibility of stitching several gratings together in order to increase the size of the field of view 34 . Tiling of the gratings as reported in the literature would allow to construct a field of view suitable for human chest imaging, and combined with a powerful enough x-ray source this would make it possible to acquire human chest radiograms with acquisition times below 1 second. Finally, the design of the grating interferometer for human imaging has to be carefully considered. On one the hand, the scattering signal obtained at e.g. 120 kVp would be significantly less than at 35 kVp, but on the other hand, the human lung is much thicker, so the signal strength should be much higher. Finally, the size of the human alveoli 35 is approx. 200 μm, whereas those of mice are approx. 60 μm in diameter 28 . Therefore, the dark-field signal from human lungs will be weaker as described in theoretical studies [36][37][38] . All these effects have to be carefully considered for the design of an interferometer for human imaging. However, the sensitivity of the interferometer can be adjusted by adapting the grating periods and grating shape [36][37][38][39][40][41] . In general, for human lungs we can expect an increase in the alveoli size by a factor of approximately 3-4, which requires a smaller grating pitch by a factor of 3-4 to obtain the same signal strength for the object of the same sample thickness. However, a human lung is thicker than a murine one approximately by a factor of 10. Therefore, even if the mean energy is increased by a factor of 2, it can be assumed that the grating periods can be slightly larger for imaging of human lungs causing a slight increase in the system length. A total length of approximately 2 meters seems reasonable, which corresponds to the length of the typical system currently used for clinical radiography. To further investigate this issue, we are currently extending our investigations to larger animals, in order to model and estimate the performance of dark-field imaging for humans and disprove concerns with respect to beam hardening or patient variability.
The proposed technique is already a valuable tool for validation of novel therapeutic strategies, allowing for non-invasive monitoring in (dose-sensitive) small-animal lung cancer experiments. Currently, only micro-CT is utilized to this end detect lung tumors in mice, applying doses of 170-280 mGy per scan 7 . Here, concerns arise with respect to adverse radiation-induced side effects which may falsify the outcome of long-term studies 7 . Conventional x-ray radiography has not been used in small-animal experiments due to the lack in sensitivity. Based on our data, x-ray radiography could represent a superior imaging modality. Compared to the conventional radiography, x-ray radiography has a much higher sensitivity, and compared to the conventional micro-CT, it requires significantly less radiation dose at the cost of volumetric information.
As the proof of concept data presented here are merely qualitative, future work should focus on quantitative analysis of these findings. Dark-field CT can be employed to allow for quantitative comparison of transmission versus dark-field signal to validate the superiority of x-ray dark-field imaging with respect to the conventional radiography 42 . These are the first proof of concept findings on mice with heterogeneous tumor load and further studies have to be carried out in order to validate the minimum size of the nodules that can be reliably detected with x-ray dark-field radiography. Small-animal lung cancer model and animal handling. For this study, 129 S/Sv-Kras tm3Tyj /J (K-ras LA2 ) mutant mice (Jackson Laboratory, Maine, USA) were used. In this mouse model, somatic activation of a latent K-ras allele with a G12D mutation at codon 12 leads to spontaneous development of early onset lung adenomas with 100% penetrance, that eventually evolve to high-grade adenocarcinomas 43 . K-ras mutation is predominantly seen in patients with non-small-cell lung cancer accounting for 20-30% of all incidences 44 .

Materials and Methods
Mice were anesthetized using intraperitoneal injection of medetomidine (500 μg/kg), midazolam (5 mg/kg), and fentanyl (50 μg/kg) during imaging and sacrificed after imaging with an overdose of ketamine and xylazin hydrochloride solution (bela-pharm, Germany). The mice were breathing freely during the image acquisition.
Small-animal scanner. The imaging was performed using a small-animal scanner 21,45 . The design of the scanner was optimized for murine lung imaging in a number of steps. After initial simulations 46 , the signal of the ex vivo lung was validated 11,12 before the interferometer was used to image an in vivo mouse 47 . Subsequently, the signal decay was validated in good agreement with simulations and micro-bubbles 10,13 and mouse models of emphysematous 15,16 and fibrotic tissue 17 . In addition to the conventional X-ray gantry, consisting of an X-ray source (fixed anode tungsten-target MCBM 65B-50 W tube (RTW, Neuenhagen, Germany)) and a flatpanel (C9312SK-06 Hamamatsu (Hamamatsu, Japan)) detector with a GOS scintillator and 50 μm pixel size, a three grating Talbot-Lau interferometer was introduced. The interferometer consists of a source grating (G0) with a period of 10 μm and gold height of 35 μm, a binary phase grating (G1) that has a period of 3.24 μm and a nickel height of 4 μm and, finally, an analyzer grating with a period of 4.8 μm and 45 μm high gold bars. The design energy of the system is 23 keV. The distances between the source, the phase grating and the analyzer grating are 30 and 14.5 cm respectively, and the interferometer is operated in the first fractional Talbot order. A piezo positioner is used to move the source grating in order to acquire the phase stepping curve 48 . A sample bed is used to position the sample right in front of the phase grating. The scanner is equipped with breath and heart rate monitoring system and a warm air fan to keep the animals temperature constant during in vivo imaging. Besides the attenuation and dark-field images, the scanner also provides the differential phase contrast images. However, these images are not included in this study and are only presented in supplementary Figure 1 for the interested reader.

Reader-study.
A blinded reader study was performed with three radiologists (four years of professional experience), who were familiar with the concept of X-ray dark-field imaging. The evaluation of the randomized transmission and dark-field images was performed separately with a certain temporal interval so that a recall bias was avoided, e.g. the readers could not build on the previous results. All readers were allowed to freely adjust the windowing and the magnification of the images using OsiriX Dicom Viewer (Pixmeo Sarl, Bernex, Switzerland). The readers were asked to mark lung cancer and delineate the boundaries of lung nodules.
Histological preparations. Following intubation of the mice and diaphragm dissection, airways were inflated with 4% (w/v) paraformaldehyde (PFA). Later, the tracheae were knotted, the lungs were excised and transferred into 4% PFA-loaded tubes for tissue fixation overnight at 46 °C. After tissue processing and paraffin embedding, 3 μm thick sections were cut and placed on Superfrost TM Plus Adhesion Slides (Thermo Scientific). The de-paraffinized sections were stained with hematoxylin and eosin (Carl Roth, Germany) subsequently, and dehydrated respectively in consecutively grading ethanol and xylene solutions (AppliChem, Germany). Dried sections were mounted with entellan (Merck, Germany).