Development of Chinese mesh-type pediatric reference phantom series and application in dose assessment of Chinese undergoing computed tomography scanning

A software named 3D-Doctor^TM (Able Software Corp. 2012) was used to segment the contours of different organs and bones from the CT images. Firstly, CT images were imported into 3D- Doctor^TM to segment the skeletons. Since the density of skeletons was quite different from other organs, it was easy to distinguish them in CT images. The semi-automatic threshold segmentation tool in 3D-Doctor^TM was used to segment the skeletons, as shown in figure 2.

Since the density of some organs was not much different from soft tissues or other organs around them, the semi-automatic threshold segmentation function in 3D-Doctor^TM software was not suitable for the segmentation of these organs. Therefore, manual segmentation was required. Most of the organ segmentation was performed in enhanced CT images, in which the contours of the organs were more apparent than in normal CT images. The median filtering operation was performed to increase the contrast of the organ in the CT images. This improved the quality of the CT image and facilitates the subsequent manual segmentation of the organ. The organs from every CT image were segmented under the guidance of a professional pediatrician to ensure the correctness of the organ segmentation. For organs (such as lungs, liver, and spleen) whose contours were not definite enough, manual segmentation was processed. For various glands (such as the adrenal gland, pancreas, thyroid gland, and thymus), the volumes themselves were small, and their contrasts in the CT images were also inferior. It was necessary to combine the CT images with the coronal images and the sagittal images according to the anatomical structure. Considering the continuity condition, which determined the boundary contour, the organs were segmented in the transverse section and modified in the coronal and sagittal sections.

The 3D surfaces of the organs and skeletons after manual segmentation were generated by using 3D-Doctor^TM. The outer contour was checked to conform to the normal organ anatomy and the missing CT images were checked as well. The surface rendering function of the 3D-Doctor^TM can develop the contours of the different organs and skeletons into polygon-mesh (PM) models.

The segmented organs, including semi-automatic threshold segmentation and manual segmentation results, were imported into Rhinoceros^TM (Robert McNeel & Associates 2018) for further processing.

The organs, which had not been segmented in the CT images, were developed via the non-uniform rational B-spline (NURBS) method. The organs such as rib cartilage, salivary glands, and tongue were transformed from the basic geometries, which were similar to the real shape. For example, the eyeballs were established as spheres. The lens, the testicle, the prostate, and the ovaries were represented by ellipsoids. The tongue was built into a cylinder and then smoothed by edge chamfering and axial compression. The cheeks were built as an ellipsoid, which was chamfered and curved. At the same time, as the soft ribs were not displayed in the CT images, we established the soft ribs manually. Several cylinders with a smaller radius at the end close to the ribs and a larger radius at the other end were modified to represent the soft ribs. They were firstly compressed and then bent along the trend of ribs. Both ends of the cylinder were chamfered. The anatomical structures of the small intestine and colon themselves were extremely complicated. It was tough to operate if a structure identical to the anatomical shape required to be established. The method developed by (Kim et al 2018) was adopted in the present work to establish the mesh-type small intestine and colon. Firstly, the general trend lines of the small intestine and colon were established. Then, based on the trend line, the colon was constructed as a changing diameter pipe, while the small intestine was kept as a constant diameter pipe. Figure 3 showed the structures of the small intestine and colon and the structure established in our phantom. Finally, the NURBS models were converted into the polygon mesh (PM) model via the Rhinoceros^TM (Robert McNeel & Associates 2018).

Zoom In Zoom Out Reset image size

The NURBS models constructed by the 3D-Doctor^TM were composed of a large number of meshes, which were not convenient for phantom adjustment. With the 'Reduce Mesh' instruction in the Rhinoceros^TM, we reduced the number of meshes to an appropriate quantity. However, in the process of mesh deletion, the smoothness of the organ was destroyed, and the surfaces of the organs became extremely rough. Decreasing the number of PM grids to target numbers at one time would change the shape of the organs and cause errors. Therefore, it was necessary to cut down a smaller number of meshes each time and smooth the organs at the same time. After many trials for this study, the number of PM grids deleted was about 10% each time, and the smoothing coefficient was set at around 0.3. These two instructions were used interchangeably. In this way, we reduced the number of PM grids at the same time to ensure the shape and the smoothness of the organs were well maintained. Figure 4 showed the PM surfaces of the liver before (left) and after (right) the optimization by using the Rhinoceros^TM.

Zoom In Zoom Out Reset image size

The process of deleting the PM grids induced various PM grid errors. If these errors were not corrected, as the mesh type phantom was voxelized, an error would occur and the original shape of the organ could not be guaranteed. There are three common errors: the exposed mesh surface, the duplicate mesh surface, and the non-manifold mesh surface. The exposed mesh surface means holes appear on the surface. These holes can usually be filled in with the 'Fill Mesh Hole' command. The second type of error is a duplicate mesh face or an extra mesh face. The presence of such a grid can cause problems with the shape or integrity of the organ. These mesh faces need to be removed by using the 'Delete Mesh Surfaces' command. The third error is a non-manifold mesh surface. A non-manifold mesh surface means that the edge is used by multiple meshes. In this situation, the program cannot distinguish between the inside and the outside organs. Therefore, without affecting the original structure, we used the commands mentioned above, 'Delete Mesh Surfaces', 'Fill Mesh Hole' and 'Patch Single Surfaces', to solve the errors of these non-manifold meshes. The primitive models were refined to turn into high-quality models by the method mentioned above.

Each pediatric mesh-type phantom was constructed based on CT images of several pediatric patients. Therefore, it was necessary to stitch the outer skins of these children together to form a complete skin. The skin was a complete enclosure in an exported file of the software 3D-Doctor^TM. The bottom PM meshes of the single part of the skin were removed at first, the command 'Patch Single Surfaces' was then used in combination with the command 'Fill Mesh Hole' to splice the different parts of the skin.

Considering that the patients' physical data was not entirely consistent with the reference data, it was required to adjust the mass of each organ to fit the reference data. The reference volume of each organ was calculated referring to the ICRU Report 46 (ICRU 1992) and GBZ/T 200.2-2007 (Ministry of Health P.R. China 2007b). The reference masses of Chinese children's organs were mainly found in the report of China's occupational health standard GBZ/T 200.1-2007 (Ministry of Health P.R. China 2007a). For the reference masses of the organs not given in the report, the Asian reference data was used here instead (Tanaka et al 1998). At the same time, the densities of the organs were found in the ICRU Report 46 (ICRU 1992), and the reference volumes of organs were calculated.

For general organs, the offset commands in Rhinoceros^TM were used to adjust the volume of the organ to match the reference volume. A negative offset represents expansion, and a positive offset represents reduction.

For an organ containing a wall to the contents, the reference volume of the content and the reference volume of the wall were first calculated, the two values were then added together to obtain the total volume of the wall and the contents. The volume was first matched to the total volume to obtain the outer wall of the organ. The offset command was then used to generate an additional PM mesh surface as being the inner wall of the organ. The volume of the inner wall was matched to the volume of the contents. The wall of the organ was located between the outer wall and the inner wall. Inside the inner wall were the contents of the organ.

Skeletons are composed of cortical bone (CB), trabecular bone (TB), red bone marrow (RBM), yellow bone marrow (YBM), cartilage, and miscellaneous bone. RBM and bone surface in trabecular bone are radiation-sensitive organs. In this modeling process, the basic PM models of the skeleton were produced by using the same conversion procedure employed for the general organs and tissues. The site-specific non-uniform bone model was then used to divide the bone into cortical bone (CB) and spongiosa. The spongiosa was uniformly mixed by TB, RBM, YBM, and miscellaneous bone. The cartilage portion was mixed into the soft tissue in the phantom. The GBZ/T 200.2-2007 provides reference masses of different skeletal tissues. table 3 listed the mass data for the bone composition of Asian children's reference phantoms, and the reference density was also obtained from the ICRU Report 46 (ICRU 1992). Bone distribution data was obtained from the ICRP Publication 70 (ICRP 1995) and ICRP Publication 89 (ICRP 2002).

Established organs and bones were assembled into a complete phantom. The overlapping parts were adjusted appropriately. It was achieved by a subtle movement of the position and transformation of the organ shape. In Rhinoceros^TM, we set series control points on the NURBS surface to complete partial adjustment and kept the mass consistent with the reference data at the same time.

Since the established mesh-type phantoms still have some tiny surface imperfection, they cannot be directly applied to the Monte Carlo simulation. The phantoms were voxelized and converted into voxel phantoms. The THUDose software(Lu et al 2017), which was developed at Tsinghua University, was used to complete the process of voxelization. The voxel sizes of Chinese pediatric reference phantoms were listed in table 4.

THUDose (Lu et al 2017) is a Monte Carlo modeling and simulation software with user-friendly UI, which is developed at Tsinghua University based on Geant4 code. The modeling and simulation processes of this work were carried out based on THUDose.

Throughout communication with Chinese pediatricians, we learned that the LightSpeed 16 CT of GE company simulated in this article is widely used clinically in China. In addition, the data of 16 slice scanners is adopted as the current national standard data(Ministry of Health P.R. China 1995). The collimation width of this type of CT includes 5, 10, and 20 mm. Moreover, there were two types of bowtie filters for the head and body. The opening angles of the fan beam for head and body scanning were 27.5° and 55°, respectively. The distance from the source to the phantom center of this type of CT was 54 cm, and the maximum scan field was 50 cm in width. The x-rays emit a fan beam from a point and this beam passes through the flat filter, which was used to absorb the low-energy part of x-rays and then the bowtie filter. The bowtie filter was used to compensate for the difference in the thickness of the human body. It was also used to balance the amount of radiation reaching the detector, thereby improving the image quality and reducing the peripheral dose of the patient.

The x-ray source was directly obtained by energy spectrum sampling. The energy spectrums adopted by x-rays were generated using XCOMP5R software (R. Nowotny 1985)for specific parameters such as tube voltage, anode angle, plate filter material, and thickness. The tube voltages used in this study were 120, 100, and 80 kVp. The anode angle was 12.0°. The structure of the flat filter was considered in this study when generating an x-ray energy spectrum. The flat filter was set as 2.5 mm thick aluminum. The distance from the source to the x-ray energy spectrum recording position was 10 cm.

The material and shape of the bowtie filter determine the final x-ray distribution. However, the actual geometry of the bowtie filter was relatively complex, and it was difficult to obtain the geometric parameters from the CT scanner manufacturer. Therefore, the bowtie filters for head and body scanning used the simplified model, as shown in figure 5, with the material of aluminum. The dose profile was experimentally measured using an ion chamber. The geometric parameters were adjusted until the dose distributions obtained by Monte Carlo simulation were coincident with experimentally measured results(Pan et al 2014). According to the information provided by the CT production, the CT bed was modeled as part of a 91 cm internal diameter cylinder with a thickness of 4.5 mm, and the material was carbon.

Zoom In Zoom Out Reset image size

CT scanning includes helical scanning and axial scanning. Organ doses absorbed from helical scanning are related to the pitch, and they are approximately equal with organ doses obtained from the axial scanning with the same scanning parameters (Pan et al 2014). Therefore, the axial scanning type was simulated in this work to generate the organ dose database. 16 x-ray sources and filters were uniformly arranged in a circle to simulate an x-ray tube rotating one circle. This method was verified by the work of (Pan et al 2014).

This modeling of CT scanner was verified by comparison of simulated and measured results. A complete system for x-ray measurements (Unfors RaySafe X1, Sweden) were used to measure the CTDI₁₀₀ value in the air from the isocenter to the edge of the CT scanner. For the lateral dose profile, the relative errors between calculation and experimental results were less than 7%. CTDI₁₀₀ values were measured in the standard CTDI head and body phantoms by using ion chambers and it showed a good agreement with simulated results. Moreover, the modeling method was verified based on pediatric physical phantom(Pan et al 2014). For the head and body filter, the CTDI_vol was measured using a head CTDI phantom with a diameter of 16 cm and a body CTDI phantom with a diameter of 32 cm.

The Chinese pediatric reference computational phantom (CRC) and the CT scanner model established in this study were further used to calculate the organ absorbed dose database using the Monte Carlo method. The collimators were moved from the head to the feet of the phantoms, and the organ doses were calculated for once axial scanning in each position. The Monte Carlo simulations were formed for the scanning conditions with x-ray tube voltages of 120, 100 and 80 kVp, with collimators of 20, 10, and 5 mm width, with filters for head and body. Finally, the datasets including organ doses obtained from single-layer axial scanning were established. The organ doses were calculated based on the CT scan organ dose databases for each phantom. The scan ranges included the head, chest, abdominal-pelvis, and chest-abdomen-pelvis (CAP). The human anatomy positions of the scan range were determined referring to the AAPM CT scan protocol (AAPM 2010) (table 5).

The absorbed doses of most organs (organs except for skin, red bone marrow, and bone surface) were directly calculated by dividing the energy deposition in the organs by organ mass. However, for skin, red bone marrow, and bone surface, the voxel model cannot describe these organs accurately. The voxel size of the phantom was generally larger than the skin thickness, resulting in a higher simulation of skin thickness and quality. The 'equivalent mass thickness' method proposed by (Liu 2010) was used to calculate the skin dose. This method reduced the skin density so that the mass thickness of the thick skin is the same as the thickness of the real skin. When the simulation was processed, the deposition energy in the low-density thick skin was recorded, and the skin dose was obtained by dividing the actual mass of the skin. The thickness of the bone surface was only 50 μm, which could not be described in the voxel model. In this study, the 3CFs-improved method proposed by (Liu et al 2009) was used to calculate the red bone marrow dose. The red bone marrow dose (${D}_{RBM}$) in different parts was as shown below. The total red bone marrow dose was mass-weighted average of the red bone marrow doses at different sites.

$$\begin{eqnarray}&&{D}_{RBM}={D}_{SPA}\cdot \displaystyle \frac{\displaystyle \int {\left(\displaystyle \frac{{\mu }_{en}}{\rho }(E)\right)}_{RBM}\times \varphi (E)\times KS(E)\times E\times dE}{\displaystyle \int {\left(\displaystyle \frac{{\mu }_{en}}{\rho }(E)\right)}_{SPA}\times \varphi (E)\times E\times dE}\end{eqnarray} \tag{ 1 }$$

where: ${D}_{RBM}$ and ${D}_{SPA}$ are the absorbed doses of red bone marrow and spongiosa at the calculated bone site. ${\tfrac{{\mu }_{en}}{\rho }}_{RBM}$ and ${\tfrac{{\mu }_{en}}{\rho }}_{SPA}$ are the mass absorption coefficients of red bone marrow and cancellous bone, respectively. $\varphi (E)$ is photon fluence in cancellous bone, assumed to be the same as photons in the red bone marrow. $KS(E)$ is a factor that considers the dose-enhancing effect of electrons generated in trabecular bone entering red bone marrow.

Based on the database of CT scan organ doses, considering the CT scan range, scan current, time, and other parameters, data was selected from the database to calculate the organ dose of the CT anthropomorphic phantom. The organ dose calculated by Monte Carlo simulation was the dose-normalized to a single x-ray photon. The number of x-ray photons in the actual CT scan was determined by the tube current and scan time, which is difficult to be quantified. To quantitatively compare the Monte Carlo simulation value with the actual measured value, a conversion factor (CF) was shown below.

$$\begin{eqnarray}&&CF=\displaystyle \frac{CTD{I}_{100measured}}{CTD{I}_{100simulated}}\end{eqnarray} \tag{ 2 }$$

where: CTDI_{100 measured} is the CTDI₁₀₀ value per 100 mAs of CT scan air measured in the CT isocenter, mGy/100 mAs. CTDI_100simulated is the CTDI₁₀₀ per unit x-ray photon under the same CT scan parameters obtained by Monte Carlo simulation, MeV g⁻¹. The organ dose was obtained by multiplying the dose value of the Monte Carlo simulation by the CF factor. The organ absorbed dose to an axial scan was calculated as shown below.

$$\begin{eqnarray}&&{D}_{j}=\displaystyle \sum _{i={z}_{1}}^{{z}_{2}}{D}_{i,j}\times CF\times \displaystyle \frac{I\times t}{100\,{\rm{mAs}}}\end{eqnarray} \tag{ 3 }$$

where: D_j is the dose of organ j, mGy. D_{i, j} is the simulated value of the average absorbed dose of the ith axis scan to the organ j. z₁ and z₂ are respectively the starting layers and terminate layers of the organ dose database corresponding to the scan range. I is the tube current, mA. t is the scan time of one revolution of CT, s.

For the helical scan, the pitch was considered. The absorbed dose of the organ was inversely proportional to the pitch as shown below. Pitch was defined as the ratio of the distance traveled by the CT per revolution and the collimation width.

$$\begin{eqnarray}&&{D}_{j}=\displaystyle \sum _{i={z}_{1}}^{{z}_{2}}{D}_{i,j}\times \displaystyle \frac{CF}{pitch}\times \displaystyle \frac{It}{100\,{\rm{mAs}}}\end{eqnarray} \tag{ 4 }$$

In order to reconstruct the first and last images during the helical scan, the actual scan range would automatically exceed the scan range of the planned image, called over-scan. The length of the over-scan increases as the pitch or collimation width increased. For the typical 16-row multi-slice CT, the over-scan length is between 3 to 6 cm. Considering the radiation dose caused by the overscan, the method for calculating the absorbed dose of the organ is as shown below, where z_over is the number of layers covered in the database for the over-scan length.

$$\begin{eqnarray}&&{D}_{j}=\displaystyle \sum _{i={z}_{1}-{z}_{over}}^{{z}_{2}+{z}_{over}}{D}_{i,j}\times \displaystyle \frac{CF}{pitch}\times \displaystyle \frac{It}{100\,{\rm{mAs}}}\end{eqnarray} \tag{ 5 }$$

In modern multi-slice helical CT devices, automatic exposure control (AEC) has become a standard device used to reduce CT radiation dose, including tube current modulation and tube voltage modulation. At present, tube current modulation technology is mainly used to automatically adjust the tube current intensity during CT scanning according to the patient's body size and attenuation characteristics. The organ dose calculation caused by CT scan using the tube current modulation technique is shown below, where I_i is the actual tube current of the ith layer. The actual tube current value of each layer set by the tube current modulation technique can be read from the DICOM file exported by the CT scanner.

$$\begin{eqnarray}&&{D}_{j}=\displaystyle \sum _{i={z}_{1}}^{{z}_{2}}{D}_{i,j}{I}_{i}\times \displaystyle \frac{CF}{pitch}\times \displaystyle \frac{It}{100\,{\rm{mAs}}}\end{eqnarray} \tag{ 6 }$$

According to the study by (Turner et al 2010), When organ doses are normalized by CTDI_vol values, the differences across scanners become very small, with a mean of 5.2%. Therefore, the organ dose caused by other CT scanners(D_CT2) can be converted from the calculated value of the GE LightSpeed 16 CT (D_CT1), as shown below. CTDI_{vol, CT2} is the volumetric CT dose index for other CT models, and CTDI_{vol, CT1} is the volumetric CT dose index for the GE LightSpeed 16 CT under the same scan parameters. It should be noted that the above method is only been proved so far for CT systems up to 64 slices.

$$\begin{eqnarray}&&{D}_{CT2}={D}_{CT1}\times \displaystyle \frac{CTD{I}_{vol,CT2}}{CTD{I}_{vol,CT1}}\end{eqnarray} \tag{ 7 }$$

The effective doses of the patient were calculated as shown below, where w_R is the radiation weighting factor. For CT radiation emission, the x-ray radiation weighting factor is 1. The w_T is the tissue weighting factor according to the ICRP Publication 103 (ICRP 2007).

$$\begin{eqnarray}&&E=\displaystyle \displaystyle \sum _{T}{W}_{T}\displaystyle \displaystyle \sum _{R}{W}_{R}\cdot {D}_{T,R}\end{eqnarray} \tag{ 8 }$$

Both VirtualDose software (Ding et al 2015) and (Lee et al 2012) calculated the doses of different types of CT for children of different ages. The pediatric phantoms used by VirtualDose and Lee et al were UF reference pediatric hybrid phantoms. The organ dose results of 1-year-old children with the same scan range using the same CT scan parameters with the works of literature were compared here.

The works conducted by (Ding et al 2015) and (Lee et al 2012) were on Siemens SOMATOM Sensation 16, which was different from the GE Light Speed 16 simulated in this study. The organ dose was divided by the CTDI_vol under the corresponding scan conditions. It should be noted that the study of Lee et al used the head CTDI_vol for normalization of all the scan ranges. For comparative purposes, the results of the chest, Abdomen-Pelvis, and CAP scan from the study of Lee et al were multiplied by the head CTDI_vol and divided by the body CTDI_vol.

The conversion factors of ^{CTDI ₁₀₀} measured and simulated (CF) with different CT scan parameters using the head and body bowtie filter were shown in tables 6 and 7. GE LightSpeed 16 CT CTDIvol measurements per 100 mAs at different tube voltages and collimation widths were shown in table 8. CTDI₁₀₀ was measured three times and averaged, and errors were calculated as the error of the mean. The errors related to the simulated values were probability statistic errors obtained by the Monte Carlo method.

Organ absorbed doses and effective doses of phantoms for different ages were shown in tables 9–12, with the values for head, chest, abdomen-pelvis and chest-abdomen-pelvis, respectively and depicted as column graphs in figure 7. The scanning parameters were tube voltage 120 kVp, tube current-time product 100 mAs, and collimation width 10 mm. A Head filter was used in the head scan simulation, and a body filter was used for chest, Abdomen-Pelvis, and CAP scan simulation. For most organs, the organ dose of the 3-month (3 m) phantom was the highest within the phantoms series, while the organ dose of the 15-year-old phantom (15y) was the lowest. Take the dose of typical organs in each scanning range as an example to illustrate the biggest difference for different ages. For each scanning range (head, chest, abdomen-pelvis, and CAP examinations), the typical organs (brain, lung, colon, and stomach) dose of 3 m were 34.2%, 6.4%, 22.1%, and 15.5% higher than that of the 15y. This illustrated that the establishment of age-specific pediatric phantoms was necessary for accurate CT dose assessment for pediatric patients.

Zoom In Zoom Out Reset image size

The comparison and relative differences in CTDI_vol normalized organ doses with VirtualDose (Ding et al 2015) and (Lee et al 2012) at different scan ranges for the 1-year-old children calculated in this study were shown in tables 13–16.

As shown in section 4.1, the differences of dose for organs wholly covered in the scanning range were relatively small. For example, the brain dose undergoing the head scan, the lung dose undergoing chest scan, the stomach, liver, and colon dose undergoing the abdomen-pelvis scan, and most chest organs and abdominal organs dose undergoing the CAP scan. The differences in physical parameters of different anthropomorphic phantoms and the position differences of the organs in the cross-section of the human body contribute to the differences in organ doses.

For organs partially covered by the scan range, the organ dose was mainly affected by the proportion of organs covered in the scan range. For example, the position thyroid in CRC01 was upper than the 1-year-old UF phantom, causing more parts of the thyroid gland were included in the head scan. Therefore, the thyroid dose result of this study was higher than VirtualDose. The same reason caused the higher results of colon dose undergoing chest scan.

For organs that were not in the scan range, the organ dose was caused by scattered photons, which was mainly affected by the distance between the organ and the scanning boundary. The closer the organ was to the scanning boundary, the more scattering dose would be received. For different anthropomorphic phantoms, the relative positions of organs in the longitudinal distribution were different, therefore there was a difference in the distance between the organs and the scanning boundary in different phantoms. The height of the CRC01 phantom and the UF phantom used in VirtualDose and Lee's study were both 77 cm. By comparing the CRC01 phantom with UF phantom, we found that the positions of abdominal organs of the CRC01 were more upper than those of UF 1-year-old phantom. Therefore, for the head scan, the abdominal organ of this study was closer to the scanning range, and dose in this study was higher. Similarly, compared to the UF phantom, the brain of the CRC01 phantom was closer to the torso. For chest, abdomen and CAP scans, the brain of CRC01 was closer to the scan range than the UF phantom, causing higher dose results. This result illustrated that the difference in organ location was an important factor affecting the results of dose calculation. In addition, for some organs very far away from the field, the doses were so low that the relative differences were high.

The AAPM recommends age-defined pediatric protocols in CT examinations (AAPM 2010), and this requirement is also implemented in Europe. However, in clinical practice in China, adult protocols were often applied to children. In recent years, the radiation protection of pediatric patients in CT examinations is drawing increasing attention in China. Therefore, in the revision of Chinese national standards for CT examinations, it is required to promote protocols specifically for age-defined pediatric patients. This work provides a database for the formulation of national standards. In clinical practice, the organ mass and organ position of different patients vary largely, and they are also quite different from those of reference phantoms. In addition, for ages before 10 (including 10-year-old), the pediatric reference phantoms use the same phantoms with two sets of sex organs. This substitution can also affect the accuracy of the results. Therefore, the simulation results based on the reference phantoms can only be applied as a reference level for dose estimation of medical diagnosis. For more accurate dose calculation requirements, for example, in radiation therapy, patient-specific phantoms need to be developed and applied. However, developing phantoms for each patient manually is quite time-consuming and impractical. Therefore, automatic methods, such as automatic organ segmentation by applying the deep learning method, should be developed to replace the manual method.

The dose results of the bone surface and red bone marrow of this work were quite different from the results of VirtualDose and Lee. The differences could be attributed to two factors, one was the skeleton model, and the other was the method used for bone dose calculation. In the stylized models, the skeleton model was generally a uniform model. However, the skeleton mesh-type model was a bone-specific model in this study. The skeleton was composed of 19 different skeleton sites, each of which was divided into the cortical bone and cancellous bone, and the red bone marrow contents of different skeleton parts were various. In the UF pediatric mesh-type (Lee et al 2008, Lee et al 2010) phantoms used by VirtualDose (Ding et al 2015) and (Lee et al 2012), the skeletal system was divided into 35 sites. In the stylized phantom, the dose of the cancellous bone was used to approximately represent the dose of the bone surface. The 3CFs-improved method proposed by (Liu et al 2009) was used to calculate the red bone marrow dose in this study. Lee used the fluence-to-dose response functions (Johnson et al 2011) to estimate the absorbed dose to active marrow and endosteum (shallow marrow).

This study conducted the GE Light Speed 16 for MC simulation, but the Siemens SOMATOM Sensation 16 was conducted by (Ding et al 2015) and (Lee et al 2012). Although the organ doses were normalized to CTDIvol to eliminate the influence of the CT scanner on the results, this can lead to differences in the range of (2.4—8.5%) with the average difference being 5.2% as shown in (Turner et al 2010).

The results of this work were calculated by Geant4 based on GE Light Speed 16, while the MCNP was adopted by VirtualDose (Ding et al 2015) and (Lee et al 2012), and they simulated the Siemens SOMATOM Sensation 16. In the future study, we will use the same MC codes (Geant4) to calculate the results of the Chinese reference phantoms and the ICRP reference phantoms for the same CT scanner. Another limitation of this work is that developed mesh-type phantoms were converted into voxel phantoms to estimate the doses. Our team is working on modifying our phantoms to achieve dose calculation with mesh-type phantoms directly.