Fabrication of a pediatric torso phantom with multiple tissues represented using a dual nozzle thermoplastic 3D printer

Abstract Purpose To demonstrate an on‐demand and nearly automatic method for fabricating tissue‐equivalent physical anthropomorphic phantoms for imaging and dosimetry applications using a dual nozzle thermoplastic three‐dimensional (3D) printer and two types of plastic. Methods Two 3D printing plastics were investigated: (a) Normal polylactic acid (PLA) as a soft tissue simulant and (b) Iron PLA (PLA‐Fe), a composite of PLA and iron powder, as a bone simulant. The plastics and geometry of a 1‐yr‐old computational phantom were combined with a dual extrusion 3D printer to fabricate an anthropomorphic imaging phantom. The volumetric fill density of the 3D‐printed parts was varied to approximate tissues of different radiographic density using a calibration curve relating the printer infill density setting to measured CT number. As a demonstration of our method we printed a 10 cm axial cross‐section of the computational phantom’s torso at full scale. We imaged the phantom on a CT scanner and compared HU values to those of a 1‐yr‐old patient and a commercial 5‐yr‐old physical phantom. Results The phantom was printed in six parts over the course of a week. The printed phantom included 30 separate anatomical regions including soft tissue remainder, lungs (left and right), heart, esophagus, rib cage (left and right ribs 1 to 10), clavicles (left and right), scapulae (left and right), thoracic vertebrae (one solid object defining thoracic vertebrae T1 to T9). CT scanning of the phantom showed five distinct radiographic regions (heart, lung, soft tissue remainder, bone, and air cavity) despite using only two types of plastic. The 3D‐printed phantom demonstrated excellent similarity to commercially available phantoms, although key limitations in the printer and printing materials leave opportunity for improvement. Conclusion Patient‐specific anthropomorphic phantoms can be 3D printed and assembled in sections for imaging and dosimetry applications. Such phantoms will be useful for dose verification purposes when commercial phantoms are unavailable for purchase in the specific anatomies of interest.


| INTRODUCTION
Commercial anthropomorphic phantoms such as the RANDO (The Phantom Laboratory Inc, Salem, NY) and ATOM Dosimetry Verification Phantoms (CIRS Inc, Norfolk, VA) have been used for decades to study the performance of imaging systems and to assess the radiation dose received by patients undergoing medical procedures. 1 These phantoms have human-like geometry and are composed of materials designed to mimic the photon attenuation and scattering properties of tissue. The tissue-substitute materials are typically formulated as epoxy resins or polyurethanes with various additives to skew the electron density. 2,3 Early versions of the RANDO phantom incorporated a real human skeleton, so no two phantoms were exactly alike. 4 Despite the widespread use of commercial anthropomorphic phantoms, several challenges limit their applications. First, commercial anthropomorphic phantoms only come in a small variety of reference sizes which do not adequately represent all patients. For instance, the RANDO phantoms are only offered in adult male and female varieties. Similarly, the ATOM phantoms are available in adult male and female varieties, as well as four pediatric sizes (newborn, 1-yr-old, 5-yr-old, and 10-yr-old). However, for some research and clinical applications, it is desirable to have phantoms of smaller, larger, or even a patient-matched size. Second, commercial phantoms are very expensive, with purchase prices as high as~$25,000. Their high cost can be attributed to manufacturing techniques requiring craftsman-like skill that have not significantly changed since the late 1970s. Phantoms are produced to order using molding and casting methods which require long manufacturing lead timesone-of-akind productions are not cost-effective with this manufacturing model. Furthermore, a researcher may wish to customize their phantom by drilling additional holes for inserting radiation dosimeters or sources. However, the permanent modification of such an expensive phantom for one-time use is rarely an attractive option. Lastly, commercial anthropomorphic phantoms have simplified anatomy, often consisting of only three uniform materials representing soft tissue, lung, and bone. Yet, the human body is quite heterogeneous, so existing physical phantoms are not sufficient for applications requiring a high level of anatomical realism.
Given these challenges, it is no surprise that researchers have sought alternatives to physical anthropomorphic phantoms. Computational anthropomorphic phantoms coupled with Monte Carlo radiation transport simulation have proven to be a more flexible and customizable approach. Indeed, unlike their physical counterparts, computational phantoms have evolved much more rapidly. 5 Today's computational phantoms come in great variety of sizes (heights and weights), with the most advanced examples containing hundreds of segmented organs or tissuesdetail incomparable to any physical phantom available for purchase today. 6,7 Nonetheless, calculations performed using computational phantoms should be benchmarked against experimental measurements for verification purposes. In most cases, however, the dosimetry data generated using advanced computational phantoms have never been experimentally verified because an equivalent physical phantom does not exist. It is much easier to create a computational version of a physical phantom than it is to bring a computational phantom to life. 8,9 The ability to custom-fabricate anthropomorphic phantoms on-demand for research and other applications would represent a significant breakthrough in the field.
Three-dimensional (3D) printing is an additive fabrication approach which is ideally suited for creating one-of-a-kind parts, thus offering great promise as a solution for the custom fabrication of physical anthropomorphic phantoms. The technology has been used in medicine for the development of surgical guides, implants, and prosthetics since the early 1990s. 10 A variety of technologies can be used, such as stereolithography (SLA), fused-deposition modeling (FDM), selective laser sintering, binder jetting, or material jetting.
The printers can range from expensive industrial models to consumer-grade desktop printers. The consumer-grade market is currently dominated by thermoplastic extruders (FDM) and prices have dropped significantly over the past decade, bringing the technology to a wider audience and new applications.
Several groups have explored the use of 3Dprinting technology for the fabrication of anthropomorphic phantoms. Alfano et al. 11 (2003) developed the STEPBRAIN using SLA which featured multiple compartments that could be filled with liquids compatible with positron emission tomography or magnetic resonance imaging. Kim et al. 12 (2006) developed a Korean male dosimetry phantom using a combination of SLA, molding, and casting. The skeleton of the phantom was printed using a SLA resin with density similar to bone; however, there was no lung-equivalent material available, so a mold was printed for casting lungs out of urethane foam. Kiarashi et al. 13 (2015) created breast phantoms using PolyJet technology but found that there was no suitable material available to simulate fat. To overcome this limitation, only the fibroglandular tissue regions were printed, leaving the adipose regions blank to be filled in by a more appropriate material as a postprocessing step. Ehler et al. 14 (2014) used FDM technology to print a human head phantom out of acrylonitrile butadiene styrene (ABS) plastic. However, they found that ABS plastic tends to warp when printing large, solid parts. For this reason, the head was printed as a hollow shell of ABS that was subsequently filled in with a wax-based soft tissue simulantthe skull was not considered. Craft and Howell 15 (2017) created a full-scale torso phantom using FDM technology. To minimize warping, the authors used polylactic acid (PLA) plastic and a sagittal-slice design; however, the phantom did not contain bones and the lung regions were left blank. Winslow et al. 16 (2009) developed a computer-controlled milling technique that, while not technically 3Dprinting, is also noteworthy because it involved three tissue-substitute materials.
Unfortunately, their method was not fully automated. Anatomical cutouts from slabs of lung, bone, and soft tissue materials were manually assembled, glued, and then sanded to create each transverse slice of the phantoms. Collectively, these efforts have identified several barriers to progress: (a) The small build volume and slow speed of many 3D printers which inhibits the fabrication of human-size parts; (b) The limited variety of 3D printing materials for simulating tissues with mass densities ranging from 0.25 g/cm 3 (adult lung) to 1.85 g/cm 3 (bone); and (c) The need to print parts with multiple materials simultaneously. We have yet to identify a 3D printing technology which can fully overcome all these challenges; however, the technology is rapidly evolving with new printers and materials hitting the mass market every year.
In this study, we show how a relatively inexpensive desktop 3D printer can be used to print a full-scale pediatric torso phantom containing five distinct radiographic regions for computed tomography (CT) imaging applications. This is achieved by combining two different PLA plastics and an anatomical model with a dual extrusion thermoplastic 3D printer. Whereas most thermoplastics have a radiographic density similar to water (~1.0 g cm -3 ), our method takes advantage of a composite plastic containing iron for simulating bone.
At the same time, we spatially vary the infill density of the printed plastic within different anatomical regions of the phantom to achieve more realistic radiographic properties, despite only using two types of plastic. Our approach is unique compared to previously published works in that the fabrication method prints the entire phantom in one build process with very minimal postprocessing and no backfilling of material. Previous studies 15 mostly used 0.4 mm-diameter nozzles which come standard on many 3D printers; however, we found that a larger nozzle was more time efficient at printing life-size phantoms.

| MATERIALS AND METHODS
This choice did not result in any significant loss in detail for our imaging application because most clinical CT scanners produce images with pixels~1 mm in size. The open-source slicing software Slic3r 17 was used throughout this study to generate the toolpath (Gcode) files for the 3D printer. We found that this software offered the critical ability to customize print settings to a greater extent compared to Ultimaker Cura, our printer's manufacturer-branded freeware.  Table 1.

2.A | Tissue-equivalent plastics
In addition to using two materials, the radiographic density of the printed parts was controlled by varying the infill density setting in Slic3r. The infill density setting is commonly used to speed up the print time and save material by allowing one to reduce the amount of plastic printed on the interior of a part. An infill density setting of 100% produces a solid part, whereas a lower setting introduces small air gaps into the part in a user-specified infill pattern. The lowest infill density setting is 0% and results in a part which is a hollow shell with a specified wall thickness. Slic3r has several different infill patterns from which the user can select. To limit the scope of our research we focused only on the rectilinear infill pattern with the default infill angle of 45 degrees. The rectilinear infill pattern gives the printed parts an internal geometry similar to that of a parallel hole collimator, with the rectangular holes aligned with the printer build axis (Z-direction; Fig. 1).

2.B | Radiographic density calibration
Cubic blocks with side length 4 cm were 3Dprinted out of each plastic (PLA or PLA-Fe) using different infill density settings ranging between 30% and 100%. The blocks were then CT scanned to generate calibration curves relating the printer infill density settings to the average CT Hounsfield unit (HU) of a printed part. For each plastic a set of up to six blocks were printed for generating the calibration curves. The top layer of each block was assigned a thickness of 0 mm so that the rectilinear infill pattern was visible, as shown in

2.C | Preparation of pediatric torso model
A 10 cm axial cross-section of a pediatric torso was selected for 3D printing as a demonstration of our phantom fabrication method. The torso geometry used in this study was based on that of a 1-yr-old male hybrid computational phantom (height 85 cm, weight 15 kg) picked from the National Cancer Institute's library of computational phantoms. 6 The whole-body computational phantom was originally developed from CT images of a 1-yr-old patient and contains over 100 presegmented organs and tissues modeled as either nonuniform rational B-spline surfaces or a polygon surface mesh (Fig. 2). We

2.D | Phantom fabrication
While the build volume of our printer was just large enough to print the phantom as one piece, we opted to divide the torso into six smaller sections (~7 cm wide, 5 cm tall) to be printed separately as shown in Fig. 2. Cylindrical holes were added to the connecting faces of the assembly so that registration pegs could be inserted.
The choice to print the phantom in sections was prudent for several reasons. First, this choice helped to minimize the risk of wasted plastic in the event of a printing failure (although this never occurred in this study). Second, our spools of PLA filament only contained 750 g of plastic; therefore, it was necessary to add a new spool of plastic filament before building each section to avoid running out of material mid-print. Lastly, printing the phantom in small sections helped prevent warping which is a common problem for large, flat prints. 13 Warping occurs because the 3D-printed material shrinks as it cools, causing stresses to build inside the cooling part resulting in delamination from the build surface.
Each anatomical region in the torso model was assigned appropriate plastic and infill settings (infill density and wall thickness) in Slic3r to achieve as realistic radiographic properties as possible. The soft tissue and bone regions of the phantom were printed using PLA and PLA-Fe, respectively. The infill density setting for each anatomical region was selected using the calibration curves created from the printed blocks described in Section 2.B. Our target CT numbers for each anatomical region were selected to match a contrast-enhanced

2.E | Phantom verification
CT images of the 3D-printed torso phantom were acquired using the same scanner settings as described in Section 2.B. The images were compared qualitatively to that of the original 1-yr-old patient CT and to CT images of a pediatric torso phantom that was previously pur- For verification purposes, we also scanned additional PLA blocks printed with an infill density of 46% and 94%. The mean CT number recorded for these blocks was directly compared to what we observed in the lung and soft tissue remainder regions of the 3Dprinted torso.

2.F | Dose measurements in cylinders
As there is no commercial phantom with the same geometry as our 3D-printed torso phantom it is challenging to do a meaningful experimental dosimetry comparison to some already recognized standard. Landauer, Glenwood, IL). The cylinders with the OSLDs inserted were placed on the bed of the CT scanner. The OSLD stored signal was read with a microSTARii reader (Landauer, Glenwood, IL) before and after a single CT body scan (120 kVp, 250 mAs).

3.A | Radiographic density calibration
Axial and sagittal CT images of the 3D-printed blocks are shown in Fig. 3. The air gaps were not clearly visible within the resolution of the CT (pixel size 0.5859 × 0.5859 × 2.0 mm) for the blocks with infill density 90% and larger. The holes, however, were visible in the images of the blocks with infill density 70% and smaller.
Measurements of the septa thickness and inter-septa spacing can be found in Table 2. As expected, the inter-septa spacing decreased with increasing infill density. The septa thickness was 0.8 mm and was relatively constant for infill densities up to 50%.
However, a septa thickness of 0.858 mm and 1.019 mm was measured for the PLA blocks with infill densities of 70% and 90%, respectively. These results suggested that the Slic3r infill algorithm is controlling both the thickness and separation of the septa when the infill density is varied. Minor variations between the PLA and PLA-Fe septa thickness and inter-septa spacing were observed, even though the printing toolpath files (G-code) were the same except for the extrusion temperature for PLA-Fe which was 30°C cooler than for PLA.

3.B | Phantom fabrication
The six pieces of the torso phantom were printed over the course of a week (Fig. 5).

3.C | Phantom verification
CT images of the 3D-printed phantom are shown in Fig. 6 along with those of the 1-yr-old patient and CIRS 5-yr-old phantom for comparison purposes. The 1-yr-old phantom anatomy (middle row) shows a high degree of similarity to that of the 3D-printed phantom (top row) because it was used as the basis for creating the computational phantom from which the geometry for 3D printing was derived. One notable difference is in the outer body contour of the phantom; when the computational phantom was created, an adjustment was made to outer body contour to match the weight of the phantom to reference person characteristics. The positioning of the clavicles is also different; the patient has arms raised as is typical T A B L E 2 Measured properties of the three-dimensional (3D)-printed blocks using polylactic acid (PLA) and PLA-Fe  For the PLA block printed at 100% no line separation was clearly visible so septa and inter-septa spacing were difficult to measure.
F I G 4 . Calibration lines relating the printer infill density to the measured computed tomography Hounsfield unit (120 kVp scan) for the blocks printed out of each type of plastic. The blocks were printed using a 0.8 mm nozzle with 0.4 mm thick layers and a rectilinear infill pattern. The data points (error bars) represent the mean (standard deviation) of the pixel HU values recorded for a 2 cm diameter spherical ROI at the center of each block.
during CT scanning whereas the computational phantom's posture was altered to have arms at the side. Despite these systematic differences, our results demonstrate the remarkable capability of 3Dprinting to capture individualized anatomy with a high degree of fidelity. Furthermore, it can be observed that the CIRS phantom has uniform density within the lung and bone regions, whereas the 3Dprinted phantom has a significant amount of texture which more closely resembles that seen in patients. A profile of the measured HU values measured laterally through the phantom is shown in Fig. 7, which was created by averaging over a 10 mm sliding widow to reduce noise. The profile for the 5-yr-old commercial phantom (not shown) is qualitatively similar.
A quantitative comparison of CT Hounsfield units in various anatomical regions was performed and the results are shown in Table 3. The infill density settings for the various regions of the 3Dprinted phantom were selected in effort to achieve target CT numbers of 183, 60, −500, 1000 HU in the heart, soft tissue remainder, lungs, and bone regions, respectively. Analysis of the CT images of the 3D-printed phantom showed that we could achieve these target

3.D | Dose measurements in cylinders
As the microSTARii reader was not calibrated for measuring absolute dose for CT x-ray beams, the OSLD readings were normalized to that of the OSLD in the cylinder with 100% infill (solid). Figure 8 shows that the OSLD response increased nonlinearly with decreasing infill density as expected. These results demonstrate that the dose reading within a 3D-printed object can be modulated by spatially varying the infill density. The differences between our targeted (original patient CT) and

| DISCUSSION
actual CT numbers measured in our 3D-printed phantom can be explained by limitations in our calibration method. First, it should be noted that the calibration blocks were measured in-air, whereas the measurements within the phantom are affected by beam hardening and scatter to a greater extent. Second, the calibration block measurements were performed using a large spherical VOI placed at the center of the blocks and far from the walls; however, the measurement conditions in the phantom were sometimes different, particularly in the case of bone. The bone structures in the phantom were small and thin, making it hard to identify a suitable internal region large enough for averaging without interference from the solid wall perimeters. Third, the calibration curves (Fig. 4) are a function of the CT tube voltage (kVp), image resolution, and image reconstruction algorithm; these were kept constant in our work, but limit the applicability of the calibration curves we generated. Lastly, we did observe some systematic differences in the way the printer laid down plastic when printing the phantom compared to the calibration blocks. For instance, the septa thickness (inter-septa spacing) in the heart region of the 3D-printed phantom was 1.089 ± 0.034 mm (0.116 ± 0.051 mm), whereas there was no visible spacing in the solid PLA block. We compared line spacing in the toolpath files (Gcode) for the heart and solid block generated by the Slic3r software and found them to be the same; therefore, the observed differences are attributed to inconsistencies in the printer. Despite our best efforts to calibrate the printer, the printing process was not as reproducible as desired. The third-party Slic3r slicing software used in this study offered the critical ability to assign the infill density to different regions in an assembly of parts; however, it did not have the capability to directly define the infill line thickness and spacing as we would have liked. As future work, it will be important to seek ways to more reliably predict the CT number of our 3D-printed objects.
One such way might be to incorporate the grid septa directly into our 3D model. featured only two print nozzles, and printing two materials solid would not provide enough variation in radiographic density for a realistic anthropomorphic phantom; one would need at least three nozzles to print a phantom with three materials to represent lung, bone, and soft tissue. Another option might be to use a printer design which can efficiently switch between three materials. Lastly, it should be noted that we could not vary the infill density continuously in our phantom. Only one infill density setting could be assigned to each anatomical region. Future FDM printers may be able to vary materials continuously through appropriate mixing of an array of plastics.

| CONCLUSION
In this study we demonstrated how an inexpensive desktop 3D printer can be used to print a full-scale pediatric torso phantom containing five distinct radiographic regions for computed tomography (CT) imaging applications despite using only two types of plastic. While our method has key limitations, our results show that the creation of patient-specific imaging phantoms of comparable quality to commercial phantoms are possible with existing 3Dprinting technology. In principle, our methods can be improved by using a customized 3D printer, slicing software, and materials; however, such efforts were beyond the scope of this study, as our intent was to use off-the-shelf supplies. With more work along these lines of research, we expect that the ability to create patient-specific phantoms on-demand will soon become a reality, and this will have important research and clinical applications throughout the field of medical physics.

CONF LICT OF I NTEREST
The authors declare no conflict of interest.