Pseudo dual-energy CT-derived iodine mapping using single-energy CT data based on a convolution neural network

Objective: The objectives of this study are: (1) to develop a convolutional neural network model that yields pseudo high-energy CT (CTpseudo_high) from simple image processed low-energy CT (CTlow) images, and (2) to create a pseudo iodine map (IMpseudo) and pseudo virtual non-contrast (VNCpseudo) images for thoracic and abdominal regions. Methods: Eighty patients who underwent dual-energy CT (DECT) examinations were enrolled. The data obtained from 55, 5, and 20 patients were used for training, validation, and testing, respectively. The ResUnet model was used for image generation model and was trained using CTlow and high-energy CT (CThigh) images. The proposed model performance was evaluated by calculating the CT values, image noise, mean absolute errors (MAEs), and histogram intersections (HIs). Results: The mean difference in the CT values between CTpseudo_high and CThigh images were less than 6 Hounsfield unit (HU) for all evaluating patients. The image noise of CTpseudo_high was significantly lower than that of CThigh. The mean MAEs was less than 15 HU, and HIs were almost 1.000 for all the patients. The evaluation metrics of IM and VNC exhibited the same tendency as that of the comparison between CTpseudo_high and CThigh images. Conclusions: Our results indicated that the proposed model enables to obtain the DECT images and material-specific images from only single-energy CT images. Advances in knowledges: We constructed the CNN-based model which can generate pseudo DECT image and DECT-derived material-specific image using only simple image-processed CTlow images for the thoracic and abdominal regions.

BJR Open;5:20220059 BJR|Open Yuasa et al models and an indirect prediction approach based on the calculation of differential images from CT high and CT low images.Owing to their complexity, these methods may not be suitable for clinical applications.Therefore, a simplified image processing-based method is required.Li et al implemented several connected deep CNN models to obtain a synthetic high-energy CT (CT high ) image from a low-energy CT (CT low ) image of the thoracic region and evaluated the accuracy of the model by measuring the pixel value for only one patient. 13However, they evaluated the accuracy of the model by focusing only on the CT value, and the image quality of synthetic image was not evaluated.The evaluations of image quality are required to evaluate the accuracy of model in detail.
Furthermore, there are few reports regarding the generation of pseudo-DECT (DECT pseudo )-derived images such as virtual noncontrast (VNC) images and IMs using deep-learning methods.The DECT pseudo -derived images from only SECT images may help to improve the diagnostic ability (e.g., pulmonary embolism, liver tumours).
The objectives of this study are to develop a CNN-based model that yields pseudo high-energy CT (CT pseudo_high ) images based on the simple image processed CT low images and to create an VNC image and IM using CT pseudo_high and CT low images for the thoracic and abdominal regions.Furthermore, image similarity and pixel value of the synthetic images were investigated to evaluate the accuracy of proposed model.

Data acquisition and image processing
In this study, we used the paired contrast enhanced CT high and CT low images to train and evaluate the proposed model.The use of the patient data was approved by Institutional Review Board of Yamaguchi University (approval number: H2019-080).
Eighty patients who underwent contrast-enhanced CT between January 2018 and February 2022 were enrolled in the thoracic and abdominal regions, respectively.We employed the data of 55, 5, and 20 patients for training, validation, and testing, respectively.The training data consisted of 1,7070 and 1,3223 image slices of the thoracic and abdominal regions, respectively.The contrast-enhanced CT high and CT low images were acquired on a dual-source DECT scanner (SOMATOM Force; Siemens Healthineers, Forchheim, Germany).The tube voltage was 90 and Sn 150 kV (150 kV with tin filter) for the thoracic region, 100 kV and Sn 150 kV for the abdominal region.The tube current was adjusted using CT-auto exposure control.All the CT images were reconstructed with a 512 × 512 matrix and a slice thickness of 1 mm using the iterative image reconstruction method.
For the contrast agent administration, an iodine contrast agent was injected at the rate of 3 to 5 ml s −1 .The total volume of the contrast agent was adapted to the patient's body weight at 2 ml/kg.The CT images were acquired using bolus tracking method, and the starting threshold value of the region of interest (ROI) was set to 100 Hounsfield Units (HU) on the pulmonary artery for the thoracic region and abdominal aorta for the abdominal regions.The data acquisition delays were 10 and 45 s after reaching the threshold values for the thoracic and abdominal regions, respectively.
Using the commercial workstation (SyngoVIA cliant; ver.4.1, Siemens Healthineers, Forchheim, Germany), the ground truth VNC (VNC truth ) images and ground truth IM (IM truth ) were calculated with CT high and CT low images to evaluate the accuracy of the model.

Model training and evaluation
Figure 1 shows the workflow in this study.The ResUnet model was used as the image generation model.The architecture and implementation details of the ResUnet are provided in Supplementary Material 1.This study has the training and evaluation sections.The proposed model was trained using the paired CT high and CT low images in the training section.For the evaluation section, the new CT low images, which were not used for the training were input into the trained model, and CT pseudo_high images were output.The steps for model training and evaluation are as follows.
In the first step, pre-image processing for CT low was carried out to create the input data of the model.To prevent adverse effect from the non-anatomical regions such as the CT couch and patient immobilization devices, they were removed using the binary mask image of the patient's body.The binary mask image was created using 3D Slicer (ver.4.7.0;Brigham and Females's Hospital, Boston, MA, USA).Regions outside the mask were set to a CT value of −1000 Hounsfield unit (HU).The accuracy and generalization performance of the deep learning model can be improved by data augmentation. 14Therefore, simple image processing-based data augmentation was applied to the original CT low images to improve the accuracy and generalization performance of the model.The resulting data-augmented three-channel images were used as the input data.to improve the accuracy and versatility of the model.The resulting data-augmented threechannel images were used as the input data.For the extraction of edges, the Sobel filter was applied to extract the edge of the organ.For the threshold processing, the areas with CT values above 100 HU were extracted for the CT low images.The CT value for threshold processing was determined according to the previous studies 15,16 In the second step, the ResUnet model was trained on a graphic processing unit (GPU; NVIDIA GeForce GTX 1080Ti) using the training data.The model was separately trained for the thoracic and abdominal regions.The three-channel images were input into the model, and the CT pseudo_high images were obtained as outputs.The model weights were updated while minimizing the difference between the generated CT pseudo_high images and the corresponding ground-truth CT high images.
In the final step, the accuracy of the model was evaluated using the new CT low image data of 10 patients.The new CT pseudo_high images were generated from the new CT low images using the trained ResUnet model and were compared with the corresponding ground-truth CT high images.After that, the pseudo-VNC (VNC pseudo ) images and pseudo-IM (IM pseudo ) were calculated using the CT pseudo_high and CT low images using the workstation for the 10 patients of thoracic and abdominal regions.For IM pseudo , the lung and liver areas were segmented to only assess iodine distribution of the lungs and livers.These images were compared with corresponding IM truth .

Quantitative evaluation metrics
To evaluate the accuracy of the proposed model, the mean CT value, image noise, MAE, and histogram intersection (HI) were calculated to compare CT pseudo_high with the ground truth CT high images.
The mean CT value in the ROI of 10 × 10 pixels was calculated using ImageJ software. 17For the thoracic region, ROIs were manually positioned on the pulmonary artery, pulmonary vein, ascending aorta, lung, spine, and muscle.For the abdominal region, those were positioned on the abdominal aorta, portal vein, liver, spleen, pancreas, kidney, and stomach.The difference in CT values between CT high and CT pseudo_high images was calculated.First, pre-image processing for CT low was carried out to define the input data for the model.The three-channel images created by stacking the original CT low image, extracted edges, and threshold-processed CT low images, which were used as input data.Second, the ResUnet model was trained using the image data from 30 patients for the thoracic and abdominal regions.The model weights were updated while minimizing the difference between the generated CT pseudo_high images and the corresponding ground-truth CT high images.Finally, the trained model was evaluated using the new CT low image data of 10 patients.The accuracy of the model was evaluated by comparing the obtained CT pseudo_high images with the corresponding ground-truth CT high images.Similarly, the VNC pseudo and IM pseudo images were compared with the corresponding VNC truth and IM truth .

BJR|Open Yuasa et al
The image noise was defined as the standard deviation (SD) of CT values in the ROI of 10 × 10 pixels.For the thoracic region, the ROI was positioned on the uniform region of the ascending aorta.For the abdominal region, it was positioned on the uniform region of the liver parenchyma. 18,19The difference in image noise between CT high and CT pseudo_high images was calculated.
The MAE between CT high and CT pseudo_high was calculated for the thoracic and abdominal regions.Furthermore, to evaluate the similarities between images, the HI was calculated as follows: where H1 j and H2 j represent the histogram values of CT high and CT pseudo_high for bin number j, respectively. 20The bin size of the histogram is 5 HU.
Similarly, the VMC pseudo was compared with the VNC truth using the CT value, image noise, MAE, and HI.For the comparison of IM, because IM was segmented to evaluate only the iodine distribution of lung and liver, IM pseudo was compared with the IM truth using only the MAE and HI.
To assess the difference in CT value and image noise, the statistical analysis was performed with the Shapiro-Wilk and Wilcoxon rank sum tests.All p-values less than 0.05 were considered to be statistically significant.

Comparison between ground truth and pseudoimages
Figure 2a and b show examples of the CT low , CT high , and CT pseudo_high images, and the differences between the CT high and CT pseudo_high images; the presented data correspond to patient 9 of the thoracic region group and to patient 11 of the abdominal region group.For the soft tissue of the thoracic region, CT pseudo_ high images, which are comparable with the CT high images, can be generated; however, as observed in the coronal and sagittal views, some discrepancy has shown at high iodine concentration area and bone edge.In contrast, the CT pseudo_high images are in good agreement with the corresponding CT high images for the abdominal region.The image noise of CT pseudo_high images improve in comparison with that of CT high images.The image noise of CT high and CT pseudo_high images are 10.6 and 6.8 HU for the thoracic region; and 13.1 and 5.8 HU for the abdominal region, respectively.The differences in image noise are −3.9 and −7.3 HU for the thoracic and abdominal regions, respectively.The image noise is reduced significantly using the proposed model (p < 0.05).
The image noises of VNC truth and VNC pseudo images are 17.2 and 5.9 HU for the thoracic region, 12.5 and 5.7 HU for the abdominal region, respectively.The image noise is reduced significantly as well as CT images (p < 0.05) Quantitative evaluation metrics between ground truth and pseudo-images Tables 1 and 2 show the comparison results of the CT value in the CT and VNC images.For the CT value of CT images, there are no significant differences between CT high and CT pseudo_high for all tissues.The mean differences in CT values are less than 6 HU.For the CT value of the VNC images, there are no significant differences between the VNC truth and VNC pseudo images, and the mean differences in the CT values are less than 10 HU.

DISCUSSION
In this study, we developed a ResUnet-based method to yield CT pseudo_high , VNC pseudo images, and IM pseudo using only simple image processed CT low images for thoracic and abdominal regions.
Some ResUnet-based methods have been reported, including those for predicting dose of radiotherapy and for CT image generation from magnetic resonance imaging (MRI). 10,21In the present study, we have successfully constructed the ResUnet model to generate the DECT imaging using only single energy CT image.
For the quantitative evaluation of the MAE values, previous studies have reported that the MAEs between the CT images generated using the MRI or the CBCT and ground truth CT images were 18.98-84.8HU. 21,22 Although the modalities were different, the MAEs in our study were lower than those in the previous studies.The accuracy of our proposed model was improved in comparison with previous reported model.
Several researchers have implemented CNN models in which high-quality DECT images were yielded from SECT images. 12,13owever, previous studies did not perform the evaluation for the image similarity and quality of synthetic images.In this study, we evaluated the synthetic images using several image metrics and our model was comparable to another previous model.Although previous studies generated highly accurate CT high images, they required multiple deep learning models and an indirect prediction approach using differential images.However, these complicated methods may not be suitable for clinical use.Our study introduces a novel approach that utilizes simple imageprocessing techniques, such as edge detection and thresholding,    to perform data augmentation on the original images.Furthermore, the proposed methods yield direct predictions of CT high images and achieve results comparable to those of previous studies. 12,13Therefore, the proposed approach has the potential to overcome the limitations of the the previous studies.
As observed in Figure 2, the CT pseudo_high image noise reduced in comparison with that of the CT high images.Schindera et al reported that an image noise reduction rate of 43.9-63.9%does not have a significant difference in the sensitivity for tumour detection. 23For our results of CT images, the noise reduction rate was approximately 25-55% for both thoracic and abdominal regions.It is possible that the proposed model can reduce the image noise without affecting the diagnosis.This reduction of image noise also affected the shape of histogram.The discrepancies of the histogram were caused by the reduction of image noise.However, the histograms show an almost identical shape between the ground-truth and synthetic images.Thus, the proposed model may prove to be helpful for noisy CT images.
The present study also focused on obtaining VNC and IMs images.IMs has been reportedly used for surrogate images of lung perfusion and liver blood flow. 7,24The VNC image is generated by subtracting the IM from the contrast-enhanced CT image, and thus the VNC image is strongly related with IM. 25 In the quantitative evaluation of the VNC image and IM, the MAEs in the present study were also found to be lower than those in previous study. 21,22For the VNC image and IM, the HIs are close to 1.000, which indicates that the similarity between the synthetic and ground truth images are high.Figure 4 shows the examples of IMs for thoracic and abdominal groups.According to a diagnostic report from an experienced radiologist, the patient in the thoracic region has pulmonary embolism (Figure 4a); the patient in the abdominal region is diagnosed with liver metastasis, which is caused by pancreatic cancer (Figure 4b).Although the IM pseudo was not in perfect agreement with IM truth , the IM pseudo could predict a perfusion change as well as IM truth .Therefore, our model showed the possibility of the generation of the IM pseudo images similar to the IM truth images.
The main advantage of the proposed model is that the CT pseudo_ high images can be acquired from only the image processed CT low images.Using the proposed models, the DECT image can be generated using only a SECT image obtained by a single CT scan.Furthermore, DECT images can be obtained from singlephase CT scan data acquired in the past.Therefore, the proposed method may be valuable not only for institutions that have implemented only SECT but also for various other institutions with DECT.Hence, the proposed model could resolve several problems, including the increase in radiation dose, discrepancy in the contrast-enhancement phase, and adverse effect of the patient's movement.The proposed model also enables DECT imaging without the installation of a special DECT scanner, which is advantageous for several institutions that may not have access to a special DECT scanner.Previous studies have reported that CT examinations using a low tube voltage contribute to a reduction in radiation dose and contrast agent volume. 26,27The proposed model can obtain the DECT image only using the CT low image, thus, may contribute to reducing the required radiation dose and contrast agents.
The proposed ResUnet model has a simple architecture in comparison with other models, such as generative adversarial network. 28Once the trained ResUnet models are generated, the new CT pseudo_high image can be yielded in a few minutes using the trained model.Therefore, DECT imaging with the developed model will be useful for clinical cases.
This study has two main limitations.First, the proposed models were trained and tested using only institutional CT data.Therefore, we used only the specified energy pair (thoracic and abdominal acquisition protocols at our institution), and multicentre studies with multiple energy pairs are needed in the future.
Second, the diagnostic ability of synthetic images and the impact of image quality (e.g., image noise and artefacts) on the diagnosis could not be evaluated in this study.Therefore, in the future, further investigations are required on the effect of image quality on diagnostic performance.

CONCLUSIONS
In this study, we successfully developed a model to yield DECTpseudo image from only simple image processed CT low images.Our comparison results indicated that the proposed model enabled to generate CT pseudo_high similar to CT high images.Additionally, it was found that the image noise was reduced using the proposed model.The VNC pseudo images and IM pseudo could be generated in a manner similar to IM truth and VNC truth using the CT pseudo_high and CT low images.The results suggest that the proposed model enabled to obtain the DECT images and to provide the materialspecific images from only SECT images.

Figure 1 .
Figure1.Schematic of the training and evaluation of the developed model.First, pre-image processing for CT low was carried out to define the input data for the model.The three-channel images created by stacking the original CT low image, extracted edges, and threshold-processed CT low images, which were used as input data.Second, the ResUnet model was trained using the image data from 30 patients for the thoracic and abdominal regions.The model weights were updated while minimizing the difference between the generated CT pseudo_high images and the corresponding ground-truth CT high images.Finally, the trained model was evaluated using the new CT low image data of 10 patients.The accuracy of the model was evaluated by comparing the obtained CT pseudo_high images with the corresponding ground-truth CT high images.Similarly, the VNC pseudo and IM pseudo images were compared with the corresponding VNC truth and IM truth .

Figure 2 .
Figure 2. Examples of the CT low , CT high , and CT pseudo_high images, and the differences between the CT pseudo_high and CT high images for patient 9 of the thoracic region and patient 11 of the abdominal region.The CT low , CT high , and CT pseudo_high images are shown in a window of (50, 300) HUs.The images that show the differences are shown on a window of (−250, 250) HUs.(a) Thoracic region: the soft tissue in the CT pseudo_high images are in good agreement with the CT high images; however, some discrepancy can be seen in the high iodine concentration areas and bone edge.(b) Abdominal region: the CT pseudo_high images agree well with the CT high image.The image noise of CT pseudo_high images is reduced in comparison with that of CT high images.

Figure 3 .
Figure 3. Examples of histograms of the CT low , CT high , and CT pseudo_high images for (a) patient 9 of the thoracic group and (b) patient 11 of the abdominal group.As it can be seen in the histograms, there are some discrepancies at the peaks of around 0 HU; however, the histogram of the CT pseudo_high images is in good agreement with that of the CT high images for both regions.

Figure 4 .
Figure 4. IM truth and IM pseudo for patient 5 of the thoracic group and patient 5 of the abdominal regions.(a) For the thoracic region, IM pseudo can predict a change of perfusion as well as IM truth .(b) For the abdominal region, IM pseudo provides the tumour location and iodine distribution as well as IM truth .

Table 3
lists the MAEs and HIs in CT, VNC, and IM images for thoracic and abdominal regions.The mean MAEs for the CT and VNC images are less than 15 HU for all the patients that are being evaluated.The HIs are close to 1.000 for all the patients, and thus those results represent that the histogram of CT values of synthetic images are similar to those of ground truth images.Figure3a and bshow examples of image histograms for the 3D volume of the CT low , CT high , and CT pseudo_high images; the presented data correspond to patient 9 of the thoracic group and to patient 11 of the abdominal group.There are some discrepancies at the peaks of around 0 HU; however, the histograms of the CT pseudo_high images are in good agreement with that of CT high images for both regions.

Table 1 .
Comparison of CT values of CT and VNC images for thoracic region

Table 2 .
Comparison of CT values of CT and VNC images for abdominal region

Table 3 .
Summary of evaluating metrics in the CT image, VNC image, and IM the thoracic and abdominal regions