Investigating the feasibility of generating dual‐energy CT from one 120‐kVp CT scan: a phantom study

Abstract Introduction This study aimed to investigate the feasibility of generating pseudo dual‐energy CT (DECT) from one 120‐kVp CT by using convolutional neural network (CNN) to derive additional information for quantitative image analysis through phantom study. Methods Dual‐energy scans (80/140 kVp) and single‐energy scans (120 kVp) were performed for five calibration phantoms and two evaluation phantoms on a dual‐source DECT scanner. The calibration phantoms were used to generate training dataset for CNN optimization, while the evaluation phantoms were used to generate testing dataset. A CNN model which takes 120‐kVp images as input and creates 80/140‐kVp images as output was built, trained, and tested by using Caffe CNN platform. An in‐house software to quantify contrast enhancement and synthesize virtual monochromatic CT (VMCT) for CNN‐generated pseudo DECT was implemented and evaluated. Results The CT numbers in 80‐kVp pseudo images generated by CNN are differed from the truth by 11.57, 16.67, 13.92, 12.23, 10.69 HU for syringes filled with iodine concentration of 2.19, 4.38, 8.75, 17.5, 35 mg/ml, respectively. The corresponding results for 140‐kVp CT are 3.09, 9.10, 7.08, 9.81, 7.59 HU. The estimates of iodine concentration calculated based on the proposed method are differed from the truth by 0.104, 0.603, 0.478, 0.698, 0.795 mg/ml for syringes filled with iodine concentration of 2.19, 4.38, 8.75, 17.5, 35 mg/ml, respectively. With regards to image quality enhancement, VMCT synthesized by using pseudo DECT shows the best contrast‐to‐noise ratio at 40 keV. Conclusion In conclusion, the proposed method should be a practicable strategy for iodine quantification in contrast enhanced 120‐kVp CT without using specific scanner or scanning procedure.

image noise to optimize the contrast-to-noise ratio (CNR). Besides image quality enhancement, DECT also allows quantification of iodine concentration, which could improve lesion conspicuity due to difference in iodine content between lesions and normal parenchyma. The algorithms for DECT acquisition are unique for each CT manufacturer, so this capability is only available for some specific scanners. 4,5 Dualsource DECT scanners contain two x-ray tubes and detector arrays for simultaneous acquisition of projection data with the sources operated at different tube potentials. Fast kilovolt-switching DECT scanners allow acquisition of dual-energy data by modulating the voltage of a single x-ray generator from low to high kilovolt peaks between alternating projections. Dual-layered DECT scanners have equipped with a modified detector with two scintillation layers to receive separate high and low image data. All these proprietary techniques have posed a burden on CT system hardware, so DECT scanners are not widely available as SECT scanner. Moreover, DECT acquisition may increase the radiation dose to patients. Hence, DECT is not a routine procedure even for contrast-enhanced CT scan in our hospital. Machine learning is attracting growing interest in both academia and industry recently. Furthermore, deep learning techniques have become the de facto standard for a wide variety of computer vision problems. [6][7][8] A deep learning model learns multiple levels of representations that correspond to different levels of abstraction from the input image to perform prediction. This study aimed to investigate the feasibility of generating pseudo DECT from one 120-kVp CT by using convolutional neural network (CNN) to derive additional information for quantitative image analysis without extra CT scans through phantom study.

2.A | Calibration phantoms
A calibration phantom set which consists of an electron density phantom and additional annuluses was used to generate training dataset for CNN optimization (Fig. 1)  The elliptical, epoxy resin-based phantom houses 17 rod inserts simulating lung (inhale: 0.195 g/cc; exhale: 0.51 g/cc), adipose (0.96 g/ cc), breast (0.991 g/cc), plastic water (1.016 g/cc), muscle (1.062 g/ cc), liver (1.072 g/cc), trabecular bone (1.161 g/cc), dense bone (1.53 g/cc). The second evaluation phantom (Ephan2) shown in [ Fig. 2(b)] has the same dimensions and base material as Ephan1, but the inserts in Ephan2 are different from those in Ephan1, including 12 rod inserts simulating different tissues and five syringes filled with iodine solution.

2.C | DECT and SECT scans
All scans were performed on a dual-source DECT scanner (Somatom Definition Flash, Siemens Healthcare, Forchheim, Germany). The imaging parameters of DECT and SECT scans used in this study are shown in Table 1. Attenuation-based tube current modulation (CARE

2.E | In-house software to generate VMCT and iodine image
In the presence of iodine, VMCT created using image-based method may contain beam-hardening artifacts, 15,16 so an in-house software for realizing the projection-based method proposed by Li et al. was implemented (Fig. 4). 17 The first step in the workflow was forward projection of CT images reconstructed in mm -1 by Siddon's ray tracing algorithm to obtain low-energy projections (L) and high-energy projections (H). 18 Next, two-material decomposition was performed to estimate the equivalent thickness of basis materials. Numerous basis materials for soft and bone tissues have been suggested. 19,20 For this study, aluminum was selected for bone tissues, while acrylic was chosen for soft tissue. The equivalent thicknesses of aluminum (x A ) and acrylic (x B ) were estimated based on the following equations: where the parameters a i , b j c i , d j (i = 0-5; j = 0, 1) represent characteristics of the x-ray beam energy spectrum. In the combination step, virtual monochromatic projections were synthesized using the following equation: where  1) and (2)] were compared with those measured using a caliber.

2.F | Quantitative evaluation
The difference between real CT images (I real ) and pseudo CT images (I pseudo ) generated by CNN was quantified by using RMSE and PSNR: where V is the number of voxels within the whole image, CNR ¼ CT# À CT# BG SD BG (6) where CT# is the mean CT number of a specified material, CT# BG and SD BG are the average and standard deviation of CT numbers of tissue equivalent background material, respectively.

| DISCUSSION
Contrast material enhancement for CT has been used since the mid-1970s. Besides providing visual enhancement between a lesion and of immature microvessels. 22 Determining a scan timing to grab the right moment of maximal contrast differences between a lesion and the normal parenchyma is crucial in contrast enhanced CT. However, the optimal timing varies among patients because it is related to numerous interacting factors, such as cardiac output, venous access, renal function, hepatic cirrhosis, and so on. [23][24][25] Consequently, the reliability of DECT-derived iodine concentration for pathologic stage classification may be affected by some of the patient-related factors.
Hence, DECT scan which quantifies iodine concentration at one time point is not used in daily practice for cancer screening and staging in our hospital. For the detection of hepatocellular carcinoma (HCC), dynamic scan which acquires 120-kVp SECT images to see the enhancement in different phases is used instead. The combination of arterial phase hyperenhancement followed by portal venous phase washout appearance strongly suggests the diagnosis of HCC. 26 However, triple-phase CT is a qualitative evaluation method and relies heavily on radiologist's subjective visual assessment. CT perfusion imaging represents an important quantitative assessment method for tumor-related vascularization, which can measure the hemodynamic parameters at the capillary level, with high temporal and spatial resolution, as well as good reproducibility. 27 But the respiratory motion and high radiation dose are major limitations that need to be overcome in order for perfusion CT to be used in clinical settings.
In this work, the feasibility of using deep learning method to generate pseudo DECT based on one 120-kVp SECT scan for quantitative image analysis has been investigated through phantom study.
According to Fig. 8, CT numbers for the same iodine syringe vary with phantom size, which was also observed in estimated iodine concentrations. Nevertheless, the estimation accuracy of in-house software was comparable to that of commercial software for real DECT imaging. Due to beam hardening, a lower CT number was observed in a larger calibration phantom for the same iodine syringe. 28 This phenomenon could increase data diversity to improve CNN's generalization accuracy. As shown in Fig. 9, the RMSE between real and pseudo CT was slightly lower in Ephan1 than that in Ephan2, although the rod inserts in Ephan1 simulating inhale lung, improve image contrast but would also increase image noise. 29 For VMCT synthesized by using real DECT, the best CNR was found in 60-keV images. However, VMCT synthesized by using pseudo DECT shows the best CNR at 40 keV. Based on our results, the difference in CT number between real and pseudo CT was little, but the image noise in pseudo CT is much lower than that in real CT (see intensity profiles in Fig. 10 and 11). The difference in noise properties between real and pseudo CT propagates to the corresponding VMCT, which may explain the difference in CNR performance shown in [Figs. 12(d) and 12(e)].
Overall, the proposed method should be a practicable workflow for iodine quantification in contrast enhanced 120-kVp SECT without using specific scanner or scanning procedure.
Several limitations to this study need to be acknowledged.

| CONCLUSION
This study investigated the feasibility of generating pseudo DECT from one 120-kVp CT by using deep learning method to quantify iodine concentration and synthesize VMCT through phantom study.
Based on our results, the accuracy of iodine concentration estimated by the in-house software with CNN-generated pseudo DECT imaging was comparable to the commercial software with real DECT imaging. Moreover, the VMCT synthesized by the proposed method could provide better image contrast than 120-kVp SECT after energy optimization. In conclusion, the proposed method should be a practicable strategy for iodine quantification in contrast enhanced 120-kVp SECT without using specific scanner or scanning procedure.

ACKNOWLEDGMENT
This work was supported in part by a grant from the Kaohsiung Medical University Research Foundation (KMU-Q110003).