Low-cost 3D-printed anthropomorphic cardiac phantom, for computed tomography automatic left ventricle segmentation and volumetry - A pilot study

Introduction: Accurate cardiac left ventricle (LV) delineation is essential to CT-derived left ventricular ejection fraction (LVEF). To evaluate dose-reduction potential, an anatomically accurate heart phantom, with realistic X-ray attenuation is required. We demonstrated and tested a custom-made phantom using 3D-printing, and examined the in ﬂ uence of image noise on automatically measured LV volumes Methods: A single coronary CT angiography (CCTA) dataset was segmented and converted to Standard Tessellation Language (STL) mesh, using open-source software. A 3D-printed model, with hollow left heart chambers, was printed and cavities ﬁ lled with gelatinized contrast media. This was CT-scanned in an anthropomorphic chest phantom, at different exposure conditions. LV and “ myocardium ” noise and attenuation was measured. LV volume was automatically measured using two different methods. We calculated Spearmans ’ correlation of LV volume with noise and contrast-noise ratio respectively om 486 scans of the phantom. Source images were compared to one phantom series with similar parameters. This was done using Dice coef ﬁ cient on LV short-axis segmentations. Results: Phantom “ Myocardium ” and LV attenuation was comparable to measurements on source images. Automatic volume measurement succeeded, with mean volume deviation to patient images less than 2 ml. There was a moderate correlation of volume with CNR, and strong correlation of volume with image noise. With papillary muscles included in LV volume, the correlation was positive, but negative when excluded. Variation of volumes was lowest at 90 e 100 kVp for both methods in the 486 repeat scans. The Dice coef ﬁ cient was 0.87, indicating high overlap between the single phantom series and source scan. Cost of 3D-printer and materials was 400 and 30 Euro respectively. Conclusion: Both anatomically and radiologically the phantom mimicked the source scans closely. LV volumetry was reliably performed with automatic algorithms. Implications for practice: Patient-speci ﬁ c cardiac phantoms may be produced at minimal cost and can potentially be used for other anatomies and pathologies. This enables radiographic phantom studies without need for dedicated 3D-labs or expensive commercial phantoms. © 2022 The Author(s). Published by Elsevier Ltd on behalf of The College of Radiographers. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).


Introduction
One measure of cardiac function is the Left Ventricular Ejection Fraction (LVEF), which is derived from the end-diastolic and endsystolic volumes (EDV/ESV) of the left ventricle (LV) according to the formula LVEF ¼ EDVÀESV EDV . Thus, accurate measurement of LV volume is paramount to correct LVEF estimation. As no in-vivo reference standard exists, this measure is imaging-based, and can be an indicator of heart failure e.g. in patients undergoing potentially cardiotoxic chemotherapy. ECG-gated, contrast-enhanced Computed Tomography (CT) has been shown to correlate well with Magnetic Resonance Imaging (MRI), the gold standard for LVEF measurement. 1,2 CT has superior spatial resolution to typical MRI LVEF protocols, while temporal resolution has been steadily improving. 3 Using CT could provide an alternative for patients with contraindications to MRI, and potentially reduce waiting lists.
LV volumetry can be done manually with endocardium contours traced on multiple slices in the short axis plane. This is timeconsuming and subject to operator variability. Fully-or semiautomatic methods are preferred in clinical practice, as they improve accuracy and reduce reading time. 4e6 Such algorithms use differences in Hounsfield unit (HU) values to segment the cavities by automatic thresholding and recognition of anatomical landmarks, such as the mitral and aortic valves. The main distinction between measurement methods is whether papillary muscles are ineor excluded from volume calculations.
A potential drawback of CT is the need to image the entire cardiac cycle, resulting in substantially higher dose than coronary CT angiography (CCTA), where full dose is typically limited to a small temporal window around the phase of minimal cardiac motion. Studies have hinted at the possibility of low-dose cardiac functional evaluation with CT, 7,8 but the influence of associated noise increase on volumetric measurements has not been systematically quantified. Knowledge of any systematic correlation of these quantities is important for designing low-dose CT protocols.
The value of 3D-printed phantoms in radiology and radiography studies has been demonstrated in various settings. 9e11 They allow for detailed representation of complex anatomies, and in studies employing ionizing radiation, they are of particular value, as the dose from repeated patient exposures is ethically unacceptable. To evaluate the performance of automatic LV volumetry software, anatomical details of the LV should be accurately reproduced, especially the mitral and aortic valves, papillary muscles and ventricle wall irregularities. Of less interest in this context are the right heart chambers and coronary arteries. Furthermore, X-ray attenuation should be as close as possible to real patients to reliably quantify the impact of changing exposure and reconstruction parameters.
Recent years have seen price reductions and increased performance of 3D printers. The software required have improved in terms of user-friendliness, with many open-source alternatives available for specific purposes of medical image segmentation and 3D printing. 12e14 In this paper we demonstrated the process of producing and testing an accurate 3D-printed representation of the left heart chambers, using open-source software and low-cost 3D printing materials and printers. The anatomical similarity was compared to the real patient images. To test whether any correlation exists between LV volume measurements and image noise/ contrast, we performed measurements across a range of different exposure conditions.

Methods and materials
A single dataset for phantom production was selected from a dataset of 51 pseudonymized CCTA examinations retrieved for a related CCTA noise simulation study. These were originally identified in the PACS system of the Hospital of South-West Jutland in Esbjerg, Denmark from CCTA's performed between June 2006 and October 2020, and selected according to criteria listed below. The National Committee on Health Research Ethics provided approval for retrieval and processing of the data (reference no. 2015994), which also included permission to produce 3D-printed models.
Studies were considered suitable if all criteria below were fulfilled.
1 Attenuation difference of at least 400 HU between left ventricle and myocardium. This is well above the Society of Cardiovascular Computed Tomography (SCCT) guidelines minimum value for CCTA 15 and was chosen to make HU-based segmentation easier. 2 Lowest possible image noise (evaluated subjectively). 3 Absence of motion, respiratory, and metal artifacts. This was assessed by the first author, with more than 16 years of experience in performing and quality assuring CCTA. 4 Isotropic or near isotropic voxel resolution and a matrix of at least 512 Â 512 pixels. 5 No pathological masses in the LV. Coronary pathology was irrelevant to the segmentation task and was not evaluated.
Examples of excluded series can be found in supplementary material 1.
The selected dataset was originally acquired at 90 kVp and a reference mAs of 250, reconstructed in 0.6 mm slice thickness/ 0.4 mm increment. The end-diastolic phase (70%) was extracted, yielding 491 images in total. To further reduce image noise, images were filtered in MATLAB (Mathworks Inc., USA) using an edgepreserving median filter with a size of 3 Â 3 pixels.
Segmentation was done with the open-source software ITK-Snap. 12 The paper by Otton et al. 16 provided valuable guidance, along with manuals of the applications used. The volume was cropped to the minimal cube containing the whole heart. The heart was segmented into four different tissue classes. For each tissue, pre-segmentation HU thresholding was performed as follows: Air/ lung tissue (À1000 to À200 HU), descending aorta and spine (200e1000 HU), myocardium, epicardial fat (À200 to 400 HU), left heart chambers and ascending aorta (400e1000 HU The active contouring was inspected in real time, with multiple runs to fine-tune the HU ranges for each class. By performing segmentation in the above order, segmentation "leakage" was minimized. For the first three tissue classes, voxels were subsampled by a factor of two during the pre-segmentation step, to reduce computing time. To simplify printing and avoid leakage of contrast media, coronaries and pulmonary vessels were excluded with a manual segmentation brush. The "myocardium" tissue class was exported to Standard Tessellation Language (STL) file format directly from ITK-Snap. Further postprocessing was done with MeshMixer (AutoDesk Inc., USA). Loose fragments resulting from segmentation errors were removed with a manual 3-dimensional unwrap brush. Minor discontinuities in the mitral and aortic valves were repaired, to ensure that these remained intact in the final print. The exterior surface of the phantom was simplified by using an adaptive reduction brush. Finally the automatic mesh inspection feature was used to detect and repair errors that might prevent correct printing. To avoid support structures within the cavities, the model was divided into two halves in the longeaxis plane. Printer code generation (slicing) was accomplished with Cura 4.12 (Ultimaker BV, The Netherlands).
The Hounsfield value of contrast enhancing myocardium is dependent on contrast media dose, scan timing, myocardial viability and tube voltage. Normal values are between 80 and 130 HU at 120 kVp. 17,18 We chose to aim for somewhat higher value around 150 HU to simulate attenuation at lower tube voltages typically used in low-dose CT protocols. The linear correlation of Hounsfield values with printing material mass density has been demonstrated. 19,20 Suitable materials should thus be between 1.20 and 1.25 g/cm 3 in density, according to Solc et al. 20 The "FilamentPM PLAþ" (Plasty Mladec, Czech Republic) was found suitable, according to manufacturer datasheets. Using a Fused Deposition Modeling (FDM) 3Dprinter (Creality CR6-SE, Shenzhen Creality 3D Technology, China) the two halves were printed, with a resolution of 0.2 mm and 100% infill. After printing, the two halves were glued together. Segmentation and views of the printed phantom are illustrated in Fig. 1.
A contrast media (CM) solution of water and Omnipaque 300 mgI/ml (GE Healthcare, Norway) was mixed with household gelatin, filled in the phantom and allowed to solidify overnight in a refrigerator. The concentration was based on previous test-tube scans of varying CM concentrations, targeting an LV attenuation of 400 HU at 120 kVp.
The 3D-printed phantom was placed inside an anthropomorphic chest phantom (Lungman N1, Kyoto Kagaku, Japan). The "lung tissue" was removed to make room for the 3D-phantom. A series of retrospectively ECG-gated helical scans were performed on a Siemens Somatom FORCE (Siemens Healthineers, Germany), with three repetitions per parameter combination. A scanner-generated synthetic ECG with a stable heartrate of 60 beats per minute (BPM) was used to provide gating signals. Tube current modulation and automatic tube current selection was disabled. In total 486 image sets were generated. Detailed parameters are listed in Table 1, and scan range visualized in supplementary material 2, on a scan projection radiograph of the setup.
Images from both the source and phantom scans were loaded into Syngo.Via© VA 40 (Siemens Healthineers, Germany) "Cardiac Function" workflow, using fully automatic LV volume calculation and segmentation. Results were noted, using both blood volume (BV) mode (excluding papillary musculature but including the left ventricular outflow tract) as well as the MRI-like method of including the papillary musculature and trabeculations (geometric mode, GM).
A MATLAB script was used to generate identical ROI's at the exact same 3-dimensional position across all image series. An axial  slice was chosen where the LV was widest. ROI's were placed in the LV, avoiding solid structures, and in the "myocardium", approximately 1 cm from the LV border. Mean attenuation and noise (standard deviation) were measured. Contrast-to-Noise Ratio (CNR) between myocardium and LV was calculated as the difference in attenuation, divided by the mean noise of the two tissues. Measurements were performed on the source scan as well, with the myocardium HU measured at four different positions (see Fig. 2). Spearman rank correlation was assessed between measured volumes and image noise and CNR respectively. A table of measurements can be found in supplementary material 3.
To test whether the software was able to correlate anatomical landmarks between patient-and phantom scans, automatic coregistration in Syngo.Via was attempted on a single phantom dataset with acquisition parameters closest to the source scans and highest CNR. Using the same dataset, we reconstructed short-axis Multi Planar Reformation (MPR) series of the LV for patient and phantom scans, at identical field of views (FOV), from the mitral valve plane to the cardiac apex. Each series consisted of 186 images. LV's were segmented, using K-means segmentation with three classes (air, contrast, soft tissue). To correct for differences in orientation, we registered the phantom segmentations onto the patient segmentations in MATLAB, using similarity transformation. We calculated the three-dimensional Dice coefficient to evaluate the precision of the LV segmentations. This ranges from zero to one, with one indicating a perfect match.

Results
LV segmentation succeeded in all repeated scans, and measurements were generated automatically. Representative screenshots from postprocessing can be seen in Fig. 3. At identical exposure (250 mAs and 90 kVp) LV volume was 96.6 and 96.0 ml in phantom and patient images respectively, a 0.6% difference, using blood volume mode. Automatic registration of the phantom dataset on the real clinical scan is also illustrated in the same figure. Volume rendering images are available in supplementary material 3. Fig. 4 depicts Hounsfield values of LV and phantom "myocardium" at various tube voltages. Specifically, the phantom LV attenuation at 120 kVp was 396 HU, while "myocardium" was 169 HU. Fig. 5 comprises scatterplots of measured volumes vs. CNR and mean of LV/myocardium image noise. The LV volume mean and standard deviation across all exposures was 95.3 and 3.1 respectively for the BV mode, versus 120.7 and 8.2 ml for GM mode. In the source patient images, LV volume was 96.0 and 122.3 ml for BV and GM modes respectively. A full table of all measurements can be found in supplementary material 3.
Spearman's rho for correlation of volumes with CNR was À0.27 for the geometric mode and 0.28 for the blood volume mode. The correlation with image noise was stronger, with À0.55 for the geometric mode and 0.63 for the blood volume mode. Rho was statistically significantly different from zero in all cases (p < 0.001). Fig. 6 shows the standard deviation of measurements by kV and reconstruction parameters.
The Dice coefficient between phantom and patient short-axis LV segmentations was 0.87 indicating good correspondence of segmented volumes. Representative slices of the segmentations are illustrated in supplementary material 5.
The 3D-printer was acquired at a cost of V400, while filament costs were approximately V30 per phantom. It took three to 4 h for segmentation and mesh generation, once experience was gained. Printing the phantom required five days in total.

Discussion
We demonstrated the feasibility of 3D-printing a phantom with anatomical similarity to a clinical scan. This was reflected by the Dice coefficient, which was high, despite all steps performed between the two segmentations, and the unavoidable data loss in each step. At similar acquisition conditions, anatomical phantom and source image similarity was sufficient to allow for automatic co-registration, using standard clinical software. The measured attenuation values were representative of real patients, which is a key requirement for reliable phantom studies.
Phantom LV volumes were strongly correlated with image noise, but less with CNR. This was somewhat surprising, given the attenuation-based measurement used in CT. One possible conclusion is that dose reduction may be accomplished without trying to preserve CNR by increasing LV contrast, e.g. scanning at very low kVp or increasing contrast media flow and volume.
It was also interesting that the correlation was opposite for the two measurement methods. One possible explanation is that when smaller structures in the LV are increasingly obscured by noise, the probability of a soft tissue pixel located in the LV cavity being assigned to the contrast volume increases, thus increasing total volume in the BV mode. This confirms the importance of consistent measurement methods, both in clinical practice, and when comparing results published in literature.
Measurement uncertainty was larger using the GM mode in all cases. The reason for this is unclear, but a possible explanation is that the left ventricular outflow tract (LVOT) is not included in GM measurements. The LVOT extent is well-defined by the aortic bulb, making for theoretically more consistent segmentation, while the mitral valve hinge points used for delineation in GM mode are easily obscured by noise. This supports the manufacturers' recommendations of using the BV mode for best accuracy. Variability of volume measures was minimized in the 90e100 kVp range, across all other combinations of acquisition conditions and measurement methods. At very low voltages, excessive noise and especially beam hardening may influence results, while at higher voltages myocardium to LV contrast is reduced. This knowledge can prove useful when setting up protocols, especially using automatic kVp selection.
While the amount of CM filled in the phantom can theoretically be measured physically, both the LV and left atrium (LA) needs to be filled to properly simulate a clinical scan. As those chambers are connected, any attempt to quantify the distribution of volumes between them was not possible. Thus, no definite reference volume exists. The CM attenuation value dependency on kVp is identical to that described by Lell et al. 21 in a swine study, using the same scanner, with approximately 10% attenuation increase per 10 kVp decrease from 120 kVp reference. This can serve as a good rule of thumb for adjusting CM and kVp relative to each other.
Verrechia-Ramos et al. 22 3D-printed cardiac phantoms for LVEF measurement with PET/CT. However, these were simplified, lacking wall irregularities, valves and other anatomical details. Furthermore, attenuation properties and CT volumetry were not examined. Mørup et al. 10 demonstrated the production of realistic coronary arteries with calcified plaques from 3D-printed molds. These could be added as a separate tissue class and coronary ostia stubs modeled open for the grafting of the flexible arteries. Potentially a fully realistic (although stationary) coronary CCTA phantom could be produced to study the influence of acquisition parameters on quantitative measurements (e.g. stenosis grade) with automated analysis tools. The phantom is cheap, compared to commercially available phantoms, e.g. "Adult heart phantom" (True Phantom Solutions, Canada) costing approximately 6000 USD. Although our phantom is very simple, and the shelf-life unknown, the low cost makes technique-and/or patient-specific phantoms accessible to most researchers. This makes it possible to perform studies of hearts with varying radiological and anatomical features.
As with any approximation of human anatomy, there are limitations. Firstly, myocardium attenuation values were almost constant regardless of tube voltage. This is contrary to real myocardial tissue, perfused with iodinated contrast, where attenuation values increase due to the K-edge of iodine. In our case, we selected an attenuation value at a mid-range kVp of 90. One possible improvement could be the use of a CM doped filament, which has been tested with mixed results and requires knowledge of and access to expensive filament production equipment. 23e25 Even then, the relative uniformity of the printing material is not representative of the "anatomical noise" present in human tissue. Right-sided heart chambers were not segmented as they were isodense in the source images, due to injection of a saline chaser bolus. Despite careful filling, the formation of air bubbles could not be completely avoided, especially around complex structures, which could potentially influence measurement. We did not perform MRI of the phantom for comparison to a gold standard. The primary benefit of MRI over CT is superior temporal resolution, 26 which is irrelevant in a stationary phantom. The spatial resolution of MRI is significantly lower than CT, with slice thicknesses typically in the range from 4 to 10 mm, compared to 0.6e2 mm for CT, making for theoretically more precise CT volumetry in the absence of motion. In the literature, inter-modality systematic variation of LV volumes has been reported. 27,28 A study found that it was necessary to use other and more expensive 3D-printing techniques and materials (silicone) for phantoms with T1/T2 relaxation times necessary to reproduce realistic myocardium MRI signal. 29 Due to the influences of these factors, we considered the generalizability of such a comparison limited. The image data used were optimal in terms of image quality as described above, and results should be interpreted accordingly. If source images with e.g. motion artefacts or low LV attenuation were used instead, segmentation accuracy and accordingly anatomical details could be expected to be less accurate. 30 Lastly, the effect of motion and finite temporal resolution cannot be reproduced in our setup. Dynamic cardiac phantoms have been demonstrated, but they represent a relatively simple approximation of cardiac anatomy, 31 and the motion patterns of the heart are complex to reproduce. Our results indicate that image noise affects volume measurements, even without motion. The interaction between noise and motion is a subject for further study. The logical next step is the reproduction of a systolic phantom to measure LVEF.

Conclusion
Manufacturing of anatomically accurate and attenuationequivalent cardiac phantoms can be performed at a very low cost, using solely open-source software packages. Anatomical detail is sufficient to allow for automatic left ventricle volumetry, using commercial clinical software, opening the way for dose-reduction studies. There seems to be a relatively strong relationship between noise and measured LV volumes depending on measurement method.

Funding
This work was supported by grants from Esbjerg Fonden, Karola Jørgensens Forskningsfond at the Hospital of South-West Jutland and the research fund of the Danish Radiographers' Association (Radiograf Rådet).

Conflicts of interest
None of the authors declare any actual or potential conflicts of interest regarding this study.