Evaluation of a preclinical photon-counting CT prototype for pulmonary imaging

The purpose of this study was to investigate a preclinical spectral photon-counting CT (SPCCT) prototype compared to conventional CT for pulmonary imaging. A custom-made lung phantom, including nodules of different sizes and shapes, was scanned with a preclinical SPCCT and a conventional CT in standard and high-resolution (HR-CT) mode. Volume estimation was evaluated by linear regression. Shape similarity was evaluated with the Dice similarity coefficient. Spatial resolution was investigated via MTF for each imaging system. In-vivo rabbit lung images from the SPCCT system were subjectively reviewed. Evaluating the volume estimation, linear regression showed best results for the SPCCT compared to CT and HR-CT with a root mean squared error of 21.3 mm3, 28.5 mm3 and 26.4 mm3 for SPCCT, CT and HR-CT, respectively. The Dice similarity coefficient was superior for SPCCT throughout nodule shapes and all nodule sizes (mean, SPCCT: 0.90; CT: 0.85; HR-CT: 0.85). 10% MTF improved from 10.1 LP/cm for HR-CT to 21.7 LP/cm for SPCCT. Visual investigation of small pulmonary structures was superior for SPCCT in the animal study. In conclusion, the SPCCT prototype has the potential to improve the assessment of lung structures due to higher resolution compared to conventional CT.

present detector technology -with pixel dimensions of 0.25 mm has been introduced with a focus on pulmonary and cardiovascular applications [15][16][17] . A different detection concept, which is currently investigated for its diagnostic range, are photon-counting detectors (PCD) 18,19 . The essential advantage of a spectral photon-counting CT (SPCCT) system is that incoming x-ray photons are directly converted in electronic signals and spectrally binned by analyzing the pulse heights generated in a semiconductor detection layer 20 . Recent developments showed promising results in the areas of abdominal [21][22][23][24][25] , cardiovascular [25][26][27][28][29][30] , neurological [31][32][33] , and nanoparticle imaging 34 . In addition to those possibilities, SPCCT will offer an improved spatial resolution due to smaller detector pixel sizes compared to the current clinical standard. The influence of electronic noise is significantly reduced in the direct-converting PCDs and can be considered as eliminated for the energy levels of incoming x-ray photons 35 . Hence, the reduced pixel dimension in PCDs comes along with a lower radiation exposure compared to (a similar reduction of detector pixel size with) energy-integrating detectors.
In this study, we investigate the resolution capabilities of a preclinical SPCCT prototype compared to a conventional CT by evaluating size and shape of lung nodules in a phantom model, measuring the modulation transfer function (MTF) and demonstrating lung structure visualization in an in-vivo acquisition of a rabbit.

Materials and Methods
CT acquisition. Images were acquired with a commercial 3 rd generation 256-row clinical CT scanner (iCT, Philips Healthcare, Best, The Netherlands) and a preclinical SPCCT prototype scanner. The clinical CT scans were matched to a CTDIvol of 7 mGy. The CT was operated with 120 kVp, 107 mAs, and two different focal spot sizes: a small focal spot resulting in high-resolution CT (HR-CT) and a standard focal spot (CT). The SPCCT was operated with a step and shoot acquisition protocol with 120 kVp, 100 mAs and 1 s gantry rotation time. The x-ray exposures of the CT and the SPCCT were chosen to equalize the Air KERMA. Acquisition parameters are listed in Table 1. Images were reconstructed with standard filtered backprojection (FBP).
Spectral photon counting CT. The preclinical SPCCT scanner (Philips Healthcare, Haifa, Israel) is based on a clinical CT system (Brilliance iCT, Philips Healthcare, Haifa, Israel) providing a conventional x-ray tube and standard beam filtration but with a limited in-plane field of view of 168 mm and a z-coverage of 2.5 mm at isocenter. The scanner is equipped with hybrid multi-bin photon counting detectors, based on ChromAIX2 ASICs (application specific integrated circuit) 36 combined with cadmium zinc telluride (CZT) as sensor material. The physical pitch of the detector pixels is 500 µm × 500 µm. The projected focal spot size is 600 µm in-plane and 700 µm in the z-direction.
Lung phantom. Patient data acquired with a conventional clinical CT system were used to build a digital model of a healthy lung. A threshold was applied to binarize the CT images and to differentiate the complex lung structure from the background. Lung nodules with two different geometries were simulated and inserted in the digital model-spheres to mimic benign nodules and spheres with spikes to mimic malignant nodules (Fig. 1), similar to the FDA lung-phantom inserts 37 . The three different sphere sizes had a diameter of 3, 6 and 9 mm. A board-certified radiologist assisted in the design process of the nodules and determined the location in the model for a realistic representation. The customized lung phantom was fabricated using an additive manufacturing Nodule segmentation. Lung nodules were segmented from the reconstructed image data (CT, HR-CT, SPCCT) for the evaluation of their volume and shape. The segmentation was performed with an in-house developed tool based on a numerical computing environment (MATLAB version R2017b, MathWorks, Massachusetts, USA). In each reconstruction, the estimated center of mass of each nodule was selected. A spherical volume of interest (VOI) around the selected center of mass was extracted from the images, with diameters of d + 1.5 mm for spheres and d + 6.5 mm for spheres with spikes (with d being the nodule size). The MATLAB-internal function kmeans, which is an implementation of the k-means clustering algorithm, was used with two clusters to separate the nodule structure from the background. Then, any non-connected component to the center of mass was removed from the segmentation. For each nodule, the segmentation was repeatedly performed three times to reduce the impact of the chosen center of mass on the measurements. Reported results for volume and shape quantification are the average over the three repeated segmentations.
Nodule volume quantification. The nodule volume was determined by multiplication of the voxel count in one segmentation with the corresponding voxel size. The standard of reference was the segmentation performed on the digital lung phantom. Due to the realistic placement of the nodules inside the lung, connected parts to the center of mass of the nodules were included within a certain VOI (see section Nodule segmentation). Nodule volumes were evaluated with linear regression analysis and by comparing the different modalities to the standard of reference in a Bland-Altman plot.
Nodule shape quantification. Due to different positioning of the lung phantom during scanning the images are not registered to each other and also not to the reference image. Therefore, nodule segmentations were semi-automatically registered to the reference segmentation of the three-dimensional (3D) printing template. In a first step, each segmentation was upscaled with cubic interpolation to isotropic voxel sizes of 0.14 × 0.14 × 0.14 mm 3 . In a second step, an expert in medical image processing measured rotation angles of the segmentations with respect to the reference. The segmentations were rotated around the x-, y-and z-axes to be in the same orientation as the reference. In a final step, two-dimensional (2D) cross-correlation of the mid-slices of the segmentation was used to shift to the same position as the reference. After registration, the Dice similarity coefficient was computed to determine how well each modality can represent the reference nodules. The Dice similarity coefficient is given by  Spatial resolution. To evaluate the in-plane resolution of the CT and HR-CT images, the vendor specific phantom (Philips iCT head system, Philips Healthcare, Haifa, Israel) with a tungsten wire diameter of 200 µm was scanned at 120 kVp in the conventional CT. For the evaluation of the in-plane resolution of the SPCCT prototype scanner a comparable self-made phantom with a wire thickness of 100 µm was applied. The phantoms were aligned so that the wire was parallel to the rotation axis of the system and close to the rotation center. The resolution was evaluated quantitatively utilizing the MTF. A small region-of-interest (ROI) around the wire was reconstructed using the same reconstruction filters and processing as for the images of the lung phantom. The MTF was then determined similar to the approach by Yu et al. 38 Several image slices were averaged to reduce noise. The background was calculated as the mean of the image region excluding the wire and subtracted from the image. The resulting image was averaged radially around the wire to calculate a one-dimensional profile. A Hankel transform was applied to the one-dimensional profile to obtain the MTF 39 . The MTF was corrected for the finite size of the wire as described by Nickoloff 40 . Finally, the MTF was normalized to achieve unity at zero frequency.
In-vivo experiment. A clinical HR-CT scan, selected from the departments Picture Archiving and Communication system (PACS), was visually compared to an in-vivo SPCCT acquisition of a New Zealand white rabbit (weight: 3.7 kg). The visual appearance was assessed by one experienced radiologist (board-certified; 4 years of experience). The study was approved by the French Department of Education and Research under the reference number APAFIS#1732-2015091411181645 V3. All experiments were performed in accordance with relevant guidelines and regulations. Patient population. Institutional review board approval was obtained prior to this study. Written informed consent was waived by the institutional review board (Ethikkommision der medizinischen Fakultät, Technical University of Munich, Germany) as all patients were included retrospectively. All scans were performed exclusively for clinical use with clinical standard protocols. Figure 2 illustrates a sagittal slice of the lung phantom with a magnification of the area around the 3 mm sphere with spikes. Edges and boundaries were more prominent in images of the SPCCT compared to CT and HR-CT. Moreover, small details such as the spikes are closer in appearance to the reference. Figure 3 shows exemplary the segmentation for the 6 mm nodules. The spherical VOI is visualized in light transparent red and the segmented nodule is visualized in opaque red. Visually, one can observe that the 3D renderings from SPCCT data give the closest representation of the ground truth. However, a small blood vessel at the bottom of the sphere with spikes, indicated by a black arrow in the reference segmentation (Fig. 3A), is lost in every modality. The connection between the vessel and the peak of the bottom spike could not be identified in any modality. Hence, the vessel is not included in the segmentations of the different modalities.

Results
Volume estimation of the nodules showed an underestimation for all modalities. The linear regression gives the best results for SPCCT (slope: 0.952; intercept: −6.842 mm 3 ) compared to HR-CT (slope: 0.942; intercept: −9.208 mm 3 ) and CT (slope: 0.933; intercept: −8.622 mm 3 ) with a root mean squared error (RMSE) of 21.3 mm 3 , 26.4 mm 3 , and 28.5 mm 3 for SPCCT, HR-CT and CT, respectively (Table 2). Figure 4A shows a plot of the linear regression. The blue line for SPCCT is the closest to the diagonal line. Bland-Altman plots show a mean difference Dice similarity coefficients were consistently superior for SPCCT (mean: 0.90) compared to HR-CT and CT (both, mean: 0.85), Table 3. The two-tail paired t-test showed a significant difference (P < 0.05) between the Dice coefficients of SPCCT and the values of HR-CT and CT (Table 3). The standard deviation of the Dice coefficients from three repeated segmentations indicate no substantial difference.
The MTF measurements are reported in Fig. 5. The 50% (10%) MTF cutoff was 6.7 (10.5), 6.1 (9.8) and 11.0 (21.7) LP/cm for CT, HR-CT and SPCCT, respectively. Figure 6 shows a comparison between images acquired of an in-vivo rabbit with SPCCT and a patient acquired with HR-CT. With HR-CT, bronchi and bronchioles down to a diameter of 1.5-2 mm could be identified. An identification of smaller bronchioles (terminal, respiratory and lobular bronchioles) is not possible due to limited resolution. Some secondary pulmonary lobes and lobular arteries (1 mm in size) can be identified. With SPCCT, very small bronchioles with a diameter of below 1 mm (corresponding to a wall thickness of 0.15 mm) can be clearly identified (Fig. 6E, marked with arrow a). The branching of the dorsal bronchiole (Fig. 6E, arrow b) shows a typical separation of lobular bronchioles, suggesting that even lobular bronchioles can be visualized. Vessels to a diameter of below 0.4 mm can be identified (Fig. 6E, arrow c). Comparing images of HR-CT and SPCCT adjusted to the same size, vessels and walls of bronchioles are visualized more distinctively.
Overall, the subjective image quality of SPCCT-images is superior to HR-CT images regarding resolution and detectability of structures.

Discussion
In this work, we investigated high-resolution imaging with a preclinical SPCCT prototype for pulmonary imaging in comparison to a commercially available CT system. We showed that the higher spatial resolution of SPCCT leads to a more precise assessment of lung nodules. Moreover, the visual investigation of small pulmonary structures was superior for SPCCT in the phantom and animal study.
Pourmorteza et al. illustrated that photon-counting detector CT (PCD-CT, a synonym for SPCCT) has the potential to provide high-resolution images with lower image noise compared to conventional CT 35 . On this note, various academic-industrial research collaborations are developing and evaluating multiple photon-counting detector concepts. While basic concepts and the ultimate goal between the different platforms are similar,  individual parameters vary from concept to concept, e.g. detector pixel-size. As this is not the focus of this work, we would like to refer interested readers to the work of Willemink et al. 41 In our study, we also observed superior high-resolution capabilities of SPCCT compared to conventional CT. However, we did not compare the noise levels of the different systems, because a fair comparison of the image noise would require the same spatial resolution   Table 3. Dice similarity coefficients for each nodule and modality compared to the reference nodules. Values close to one indicate a high similarity to the reference. Dice coefficients are given as mean of three repeated segmentations with standard deviation (mean ± SD). The paired t-test suggests a significant difference between the Dice coefficients for SPCCT and conventional CT (CT, HR-CT). * Indicates a significant difference (P < 0.05). for both systems. This would imply to reduce the spatial resolution of the SPCCT images, what is not intended in this study. It is known that higher spatial resolution results in more image noise given the same radiation dose. Reduced detector pixel sizes lead to a decreased number of photons reaching each detector element. The reduced statistics at the detector generates an uptake in noise. Moreover, in SPCCT high-energy photons and low-energy photons are weighted to contribute equally to the signal. In contrast, in conventional CT high-energy photons contribute relatively more to the signal than low-energy photons resulting in an uptake in image noise 41 . When it comes to low-dose CT another effect contributes to an increased noise level. The contribution of electronic detector noise increases in conventional CT. SPCCT, on the other hand, eliminates electronic detector noise to a certain extent by counting the photons resulting in lower image noise at same resolution. There were several limitations of this work. We used FBP instead of advanced iterative reconstruction. Iterative reconstruction is known to deliver improved image quality compared to traditional FBP and could probably improve the results for both scanners, the clinical CT 42 and the SPCCT 43 . However, due to regularization and other non-linearities, the evaluation of resolution becomes more challenging with iterative reconstruction 44 . With FBP, on the contrary, a more suitable comparison between conventional CT and SPCCT is feasible because effects of the reconstruction algorithms are reduced. Another limitation is the uncertainty in the production process of the lung phantom. Synthetic lung nodules were defined in the digital human lung model. For the 3D printing process, the digital lung model was used as input to the printer. During these processing steps, as well as during 3D printing, small errors might be propagated to the phantom due to interpolation or manufacturing processes. This might partly contribute to the discrepancy between the measured and the reference volumes. However, the RMSE in this work (21.3-28.5 mm 3 ) is in the same range as reported by Zhou et al. in a similar study assessing lung nodules (21.6-28.3 mm 3 ) 45 .
The presented results give a promising outlook to the high-resolution capabilities of the SPCCT prototype for pulmonary imaging. Higher spatial resolution, better assessment of lung nodule volume, and improved visibility of lung vessels compared to conventional CT and HR-CT were achieved. This would not only allow an earlier detection and more precise classifications of lung nodules but also improve the diagnostic confidence of radiologists assessing other pulmonary abnormalities, like COPD. In conclusion, the assessment of lung nodules could be improved with the presented preclinical SPCCT prototype. Especially the investigation of small pulmonary structures is improved due to higher resolution and the subjective higher image quality.

Data Availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.