Abstract
Purpose:
Current standard practice for clinical radionuclide dosimetry utilizes reference phantoms, where defined organ dimensions represent population averages for a given sex and age. Greater phantom personalization would support more accurate dose estimations and personalized dosimetry. Tailoring phantoms is traditionally accomplished using operator-intensive organ-level segmentation of anatomic images. Modern mesh phantoms provide enhanced anatomical realism, which has motivated their integration within Monte Carlo codes. Here, we present an automatable strategy for generating patient specific phantoms/dosimetry using intensity-based deformable image registration between mesh reference phantoms and patient CT images. This work demonstrates a proof-of-concept personalized dosimetry workflow, presented in comparison to the manual segmentation approach.
Methods:
A linear attenuation coefficient phantom was generated by resampling the PSRK-Man reference phantom onto a voxel grid and defining organ regions with corresponding Hounsfield unit (HU) reference values. The HU phantom was co-registered with a patient CT scan using Plastimatch B-spline deformable registration. In parallel, major organs were manually contoured to generate a ‘ground truth’ patient-specific phantom for comparisons. Monte Carlo derived S-values, which support nuclear medicine dosimetry, were calculated using both approaches and compared.
Results:
Application of the derived B-spline transform to the polygon vertices comprising the PSRK-Man yielded a deformed variant more closely matching the patient’s body contour and most organ volumes as-evaluated by Hausdorff distance and Dice metrics. S-values computed for fluorine-18 for the deformed phantom using the Particle and Heavy Ion Transport code System showed improved agreement with those derived from the patient-specific analog.
Conclusions:
Deformable registration techniques can be used to create a personalized phantom and better support patient-specific dosimetry. This method is shown to be easier and faster than manual segmentation. Our study is limited to a proof-of-concept scope, but demonstrates that integration of personalized phantoms into clinical dosimetry workflows can reasonably be achieved when anatomical images (CT) are available.
INTRODUCTION
In nuclear medicine, human computational phantoms comprise a digital representation of patient anatomy for radiation dose calculations, and can be broadly classified in order of increasing degrees of personalization, within the following morphometric categories:
Reference phantoms, where patient matching is by age/gender only
Patient-dependent phantoms, where matching is by age/gender and height/weight
Patient-morphed phantoms, where matching is by age/gender, height/weight, and posture/body contour
Patient-specific phantoms, where matching is uniquely defined by tomographic imaging
While bespoke patient-specific phantoms would ideally be used for all personalized dose estimation, their use in clinical workflows is sparse due to the time-consuming manual segmentation required to generate them, particularly when fine or complex structures (e.g. intestine wall) must be uniquely defined as source or target regions. For this reason, clinical dosimetry is overwhelmingly derived from reference phantoms, even with the knowledge that error in dosimetry due to organ mismatch is substantial, 20%-60%(1). In recent years, we have seen more focus and development of research using more personalized patient-dependent phantoms. These phantoms are typically generated with mesh manipulation tools within computer-aided design software (e.g. Rhinoceros, Blender) with a relatively large degree of manual input(2). These workflows generally take limited or no advantage of the relatively ubiquitous CT anatomical images that accompany most PET/nuclear medicine procedures.
In this work, we demonstrate the use of automated CT intensity-based deformable image registration to generate patient-morphed phantoms from pre-existing reference phantoms. We show that for many organs, the anatomical accuracy of the morphed phantoms, in the context of patient absorbed dose calculations, can approach that of patient specific phantoms, but generated with much less operator interaction. This effort supports a bigger picture advancement of the field towards user-friendly tools for personalized dosimetry estimation.
METHODS
Description of patient, patient CT acquisition, and patient-specific phantom construction
A sodium fluoride (18F-NaF) whole-body positron emission tomography/x-ray computed tomography (PET/CT) scan was selected randomly from the NaF-PROSTATE collection of The Cancer Imaging Archive(3) (TCIA) database (#0007), with the stipulation that arms down (AD) positioning was utilized. NaF is a oncological PET tracer with characteristic uptake throughout the skeletal system and predominantly renal clearance. The patient is male, height 1.85 m, 82.2 kg. The CT scan was acquired without contrast enhancement. For the PET modality, 122 MBq of [18F]-sodium fluoride was administered intravenously and the patient imaged approximately 1h post-administration. We segmented major organs (heart, lungs, liver, gallbladder, spleen, pancreas, stomach, kidneys, bone, brain, urinary bladder, and eyes) using a combination of manual and semi-automatic segmentation techniques using 3D Slicer (www.slicer.org) including manual segment delineation, contour interpolation, flood filling, thresholding, masking, segment smoothing, and Boolean operations. With the exception of the urinary bladder which was segmented based on PET, all segmentation was performed based on the CT anatomical reference. A ‘ground truth’ patient specific phantom was generated from the segmentation following export of the segmentation as a surface mesh, which was also converted to its corresponding tetrahedral mesh representation via Delaunay tetrahedralization with Tetgen(4) using the -pAYk flag. Ground truth organ volumes were calculated as the sum of volumes of tetrahedral elements comprising each organ in this phantom using an in-house python script.
Image registration
The tetrahedral reference Korean man phantom (THRK-Man)(5) and its polygonal surface mesh analog (PSRK-Man) are equivalent with respect to their bulk geometry and are derivable from one another; either form was used as-described for convenience unless otherwise stated. The tetrahedral volume elements of the THRK-Man are defined by identification labels (integers corresponding to the specific organ regions) which were resampled into a voxel grid with isotropic 10mm resolution using the geometry output mode of the Particle and Heavy Ion Transport code System(6) (PHITS) through PARaDIM (www.paradim-dose.org)(7). In the output volume, each organ ID was replaced with a corresponding reference value(8–10) for mass attenuation coefficient quantified in Hounsfield units (i.e. an attenuation coefficient phantom(11) for non-contrast enhanced CT); we refer to this volume as the ‘HU phantom’. The HU phantom, the patient CT image, and the PSRK-Man phantom were imported directly into 3D Slicer.
Prior to registration, the legs were excluded by cropping the HU phantom just below the inguinal region. The HU phantom was then co-registered with the patient CT image by automatic rigid registration, and the resulting transform was applied to the PSRK-Man and ‘hardened’ (on both PSRK-Man and the HU phantom). The rigidly co-registered PSRK-Man was saved in order to compare with its deformed representation (vide infra) via Hausdorff distance and Dice metrics.
Subsequently, deformable registration as-implemented via the Plastimatch(12) extension of 3D Slicer (www.slicer.org) (13), was used to morph the rigidly co-registered HU phantom to the patient CT image in a single deformation stage. As the HU phantom was sampled over reference CT HU values and thus over a similar contrast/intensity range as the patient CT, the mean squared error cost function was utilized (as-implemented in Plastimatch, this metric is intended for co-registration of CT to CT). The regularization metric in 3D Slicer’s implementation of Plastimatch balances image intensity matching with vector field smoothness; the default value of 0.005 tended to produce sharper deformations of some of the phantom organs that were deemed unrealistic based on visual assessment; thus, this metric was increased to 0.02 to lessen this effect, in combination with utilization of 150 mm isotropic B-spline grid spacing and 2×2×1 subsampling rate. 50 maximum iterations (default) were allowed, and the algorithm required approximately 1 minute of computation time under GPU hardware specification (NVIDIA GeForce GTX 1080). The resulting B-spline transform was then applied to the PSRK-Man. For S-value calculations (vide infra) the deformed surface-mesh was converted to tetrahedral format with Tetgen.
Method validation
The accuracy of both the rigid and rigid + deformable registration methods was evaluated by comparison of organ volumes, maximum, mean, and 95th percentile Hausdorff distances(14,15), and Dice similarity coefficients(14,16), between the surface mesh representations of major organs of the respective phantoms, relative to the patient specific approach. Organs with separate contents and wall components (e.g. stomach, heart, bladder) were treated as uniform organs for evaluation of agreement, as the CT contrast and/or resolution was insufficient to reliably segment the wall/contents separately; to accomplish this, relevant surfaces were removed from the PSRK-Man phantom with the computer-aided design software Blender. All segmentation and analysis was performed with 3D Slicer. Additionally, MIRD S-values(17) for the clinical workhorse PET radionuclide fluorine-18 were calculated for each phantom with PARaDIM(7).
Statistical analysis
Prism 7 (GraphPad Software, San Diego, CA) was used to perform all statistical analyses. Differences were evaluated using a two-tailed parametric paired Student’s t test, with a p value of less than 0.05 being considered as statistically significantly different. All data are presented as mean ± SD, where applicable.
RESULTS
Relative to those of the rigidly co-registered reference phantom, most organs of the deformed phantom showed improved agreement with those segmented directly, based on visual assessment of overall body contour (note chest, waist, skull), posture (note elevated head, angle of flexion of right arm), and most organ surfaces (Figure 1). Organ boundary agreement evaluated based on the mean and 95th percentile Hausdorff distance was better (i.e. smaller distance) for all organs, while all organs except for the lungs, brain and skeleton showed better agreement based on the maximum Hausdorff distance (Figure 2). Similarly, for the deformed phantom, organ overlap agreement based on the Dice similarity coefficient was equivalent or better (i.e. higher) for all organs except the brain (which displayed negligible difference). Mean values for each metric, taken over all structures of the respective phantoms, were all significantly improved for the deformed phantom (p < 0.05; Figure 2D,E). For all organs except the skeleton and brain (which was negligibly different), centroid positioning error measured relative to the directly segmented phantom was decreased for the deformed phantom. Closest-point distances between the mesh vertices comprising the reference/deformed phantoms, versus the direct segmentation-derived phantom, showed overall qualitatively improved agreement (Figure 3).
Figure 1:
Phantoms derived from PET/CT segmentation, and morphing of reference phantom via deformable image registration. A) Volume-rendered CT and exterior surfaces of segmented phantom (red), rigidly co-registered reference phantom (blue), and deformably co-registered reference phantom (green). Note the improved match of exterior body contour evident by visual assessment, of the deformed phantom relative to the original (rigidly co-registered) phantom. B) Volume-rendered CT windowed for bone contrast; interior organs of each phantom. C) Phantom surface cross-section superimposed on coronal (left), sagittal (middle) and axial (right) CT slices. Arrows represent deformation vectors applied to reference phantom.
Figure 2:
Phantoms derived from PET/CT segmentation (i.e. patient-specific phantom), rigid registration of reference phantom, and morphing of reference phantom via deformable image registration, compared quantitatively by A) organ volume, B) Dice similarity coefficient, and C) Maximum, mean, and 95th percentile Hausdorff distances. Hausdorff and Dice metrics are taken relative to the direct segmentation-derived phantom. D,E) Summary statistics (all phantom components) for Hausdorff and Dice metrics between each method. p-values apply to two-tailed parametric paired Student’s t tests.
Figure 3:
Closest-point distances between vertices of deformable-/rigidly-registered mesh phantoms and the direct segmentation-derived phantom. Negative (blue) distances indicate organ surfaces that are more central relative to corresponding organs of the segmentation-derived phantom; positive (red) distances indicate surfaces that are more peripheral.
‘Cross’ irradiation S-values (i.e. those for which the source and target regions are different) and ‘self’ S-values (i.e. for which source and target regions are identical) computed with PARaDIM for the reference phantom showed a wide range of relative deviation in comparison to those computed for the segmentation-derived analog (Table 1; Figure 4 A, C), but generally were in the expected range of 20-60%(1) with a few large deviations. Both cross- and self S-values for organs of the deformed phantom showed modest improvement in overall agreement (as compared with the reference phantom relative to the segmentation-derived phantom)(Table 1; Figure 4 B, C).
Table 1:
S-values for 18F quantified in mGy/(MBq·h) for each phantom.
| Source organ | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Target organ | Phantom | Bladder | Kidneys | Gallbladder | Liver | Pancreas | Stomach | Heart | Lungs | Brain | Spleen | Bone | RST |
| Bladder | Patient-specific | 4.92E-01 | 9.35E-04 | 8.81E-04 | 4.64E-04 | 7.67E-04 | 3.11E-04 | 1.14E-04 | 1.09E-04 | 3.77E-06 | 2.66E-04 | 3.82E-03 | 3.52E-03 |
| Reference | 3.49E+00 | 1.10E-03 | 1.14E-03 | 3.60E-04 | 9.94E-04 | 7.60E-04 | 1.10E-04 | 1.50E-04 | 1.01E-06 | 6.44E-04 | 3.94E-03 | 3.97E-03 | |
| Patient-morphed | 2.54E+00 | 1.40E-03 | 1.19E-03 | 3.57E-04 | 1.03E-03 | 8.83E-04 | 8.79E-05 | 1.65E-04 | 2.39E-06 | 8.47E-04 | 3.72E-03 | 3.87E-03 | |
| Kidneys | Patient-specific | 9.19E-04 | 3.81E-01 | 8.09E-03 | 9.24E-03 | 1.34E-02 | 3.11E-03 | 2.10E-03 | 1.92E-03 | 5.24E-05 | 1.41E-02 | 2.35E-03 | 3.41E-03 |
| Reference | 1.11E-03 | 5.20E-01 | 1.11E-02 | 5.05E-03 | 1.28E-02 | 5.72E-03 | 1.30E-03 | 2.40E-03 | 2.45E-05 | 1.21E-02 | 2.72E-03 | 3.84E-03 | |
| Patient-morphed | 1.38E-03 | 3.88E-01 | 9.01E-03 | 3.45E-03 | 9.53E-03 | 4.04E-03 | 8.71E-04 | 1.74E-03 | 3.04E-05 | 8.82E-03 | 2.25E-03 | 3.34E-03 | |
| Gallbladder | Patient-specific | 8.88E-04 | 8.02E-03 | 4.25E+00 | 2.44E-02 | 1.59E-02 | 5.65E-03 | 1.46E-03 | 1.34E-03 | 3.93E-05 | 1.04E-03 | 1.13E-03 | 2.97E-03 |
| Reference | 1.10E-03 | 1.12E-02 | 1.90E+00 | 1.88E-02 | 2.87E-02 | 1.78E-02 | 1.52E-03 | 2.11E-03 | 1.64E-05 | 3.95E-03 | 1.84E-03 | 3.50E-03 | |
| Patient-morphed | 1.14E-03 | 9.05E-03 | 9.85E-01 | 1.40E-02 | 2.02E-02 | 1.39E-02 | 1.21E-03 | 1.53E-03 | 3.33E-05 | 2.71E-03 | 1.48E-03 | 2.99E-03 | |
| Liver | Patient-specific | 4.67E-04 | 9.23E-03 | 2.48E-02 | 1.18E-01 | 1.29E-02 | 6.73E-03 | 4.87E-03 | 5.05E-03 | 8.91E-05 | 2.01E-03 | 1.61E-03 | 2.39E-03 |
| Reference | 3.71E-04 | 5.02E-03 | 1.89E-02 | 1.49E-01 | 1.16E-02 | 1.36E-02 | 7.93E-03 | 5.45E-03 | 5.22E-05 | 5.15E-03 | 1.68E-03 | 2.81E-03 | |
| Patient-morphed | 3.54E-04 | 3.46E-03 | 1.40E-02 | 1.02E-01 | 9.03E-03 | 1.05E-02 | 7.14E-03 | 4.66E-03 | 8.85E-05 | 3.13E-03 | 1.46E-03 | 2.45E-03 | |
| Pancreas | Patient-specific | 7.96E-04 | 1.35E-02 | 1.60E-02 | 1.31E-02 | 1.28E+00 | 1.46E-02 | 5.26E-03 | 2.83E-03 | 5.28E-05 | 3.89E-03 | 2.03E-03 | 3.65E-03 |
| Reference | 9.81E-04 | 1.28E-02 | 2.90E-02 | 1.17E-02 | 1.31E+00 | 4.56E-02 | 2.72E-03 | 2.67E-03 | 2.35E-05 | 2.51E-02 | 2.42E-03 | 3.81E-03 | |
| Patient-morphed | 1.00E-03 | 9.61E-03 | 2.03E-02 | 9.07E-03 | 7.27E-01 | 2.94E-02 | 2.44E-03 | 2.62E-03 | 4.31E-05 | 1.68E-02 | 2.28E-03 | 3.37E-03 | |
| Stomach | Patient-specific | 3.20E-04 | 3.14E-03 | 5.76E-03 | 6.78E-03 | 1.47E-02 | 1.99E-01 | 1.00E-02 | 4.10E-03 | 1.17E-04 | 3.84E-03 | 1.12E-03 | 2.68E-03 |
| Reference | 7.45E-04 | 5.90E-03 | 1.80E-02 | 1.38E-02 | 4.53E-02 | 5.18E-01 | 3.94E-03 | 2.92E-03 | 2.60E-05 | 2.20E-02 | 1.57E-03 | 3.19E-03 | |
| Patient-morphed | 8.43E-04 | 4.11E-03 | 1.41E-02 | 1.06E-02 | 2.95E-02 | 2.75E-01 | 3.60E-03 | 2.82E-03 | 4.59E-05 | 1.32E-02 | 1.48E-03 | 2.94E-03 | |
| Heart | Patient-specific | 1.19E-04 | 2.10E-03 | 1.46E-03 | 4.84E-03 | 5.25E-03 | 9.99E-03 | 2.57E-01 | 1.32E-02 | 2.50E-04 | 4.59E-03 | 2.22E-03 | 2.45E-03 |
| Reference | 1.07E-04 | 1.30E-03 | 1.61E-03 | 7.94E-03 | 2.72E-03 | 3.92E-03 | 2.55E-01 | 1.08E-02 | 1.39E-04 | 3.05E-03 | 1.94E-03 | 2.84E-03 | |
| Patient-morphed | 1.03E-04 | 8.71E-04 | 1.21E-03 | 7.19E-03 | 2.44E-03 | 3.56E-03 | 2.19E-01 | 1.04E-02 | 2.57E-04 | 2.11E-03 | 1.90E-03 | 2.54E-03 | |
| Lungs | Patient-specific | 9.64E-05 | 1.91E-03 | 1.33E-03 | 5.06E-03 | 2.82E-03 | 4.07E-03 | 1.32E-02 | 1.68E-01 | 3.81E-04 | 6.20E-03 | 2.92E-03 | 2.70E-03 |
| Reference | 1.37E-04 | 2.38E-03 | 2.10E-03 | 5.49E-03 | 2.70E-03 | 2.84E-03 | 1.08E-02 | 1.44E-01 | 1.97E-04 | 3.79E-03 | 2.57E-03 | 3.35E-03 | |
| Patient-morphed | 1.45E-04 | 1.74E-03 | 1.53E-03 | 4.68E-03 | 2.59E-03 | 2.74E-03 | 1.04E-02 | 1.32E-01 | 3.14E-04 | 3.65E-03 | 2.26E-03 | 2.93E-03 | |
| Brain | Patient-specific | 2.82E-06 | 4.17E-05 | 3.74E-05 | 8.57E-05 | 4.77E-05 | 1.05E-04 | 2.34E-04 | 3.76E-04 | 1.68E-01 | 1.32E-04 | 4.33E-03 | 7.59E-04 |
| Reference | 1.22E-06 | 2.55E-05 | 2.18E-05 | 5.24E-05 | 2.20E-05 | 2.26E-05 | 1.27E-04 | 2.03E-04 | 1.46E-01 | 2.96E-05 | 3.69E-03 | 1.25E-03 | |
| Patient-morphed | 2.55E-06 | 2.84E-05 | 2.50E-05 | 8.34E-05 | 3.16E-05 | 3.60E-05 | 2.45E-04 | 3.02E-04 | 1.68E-01 | 4.18E-05 | 3.33E-03 | 1.06E-03 | |
| Spleen | Patient-specific | 2.49E-04 | 1.42E-02 | 1.10E-03 | 1.99E-03 | 3.85E-03 | 3.78E-03 | 4.63E-03 | 6.21E-03 | 1.38E-04 | 7.73E-01 | 2.35E-03 | 3.19E-03 |
| Reference | 6.62E-04 | 1.21E-02 | 3.94E-03 | 5.17E-03 | 2.52E-02 | 2.18E-02 | 3.01E-03 | 3.80E-03 | 3.08E-05 | 1.04E+00 | 1.71E-03 | 3.61E-03 | |
| Patient-morphed | 9.30E-04 | 8.89E-03 | 2.72E-03 | 3.15E-03 | 1.67E-02 | 1.30E-02 | 2.08E-03 | 3.62E-03 | 4.48E-05 | 5.77E-01 | 1.62E-03 | 3.32E-03 | |
| Bone | Patient-specific | 3.94E-03 | 2.44E-03 | 1.23E-03 | 1.71E-03 | 2.11E-03 | 1.19E-03 | 2.30E-03 | 2.97E-03 | 4.36E-03 | 2.40E-03 | 2.56E-02 | 2.63E-03 |
| Reference | 3.97E-03 | 2.81E-03 | 1.98E-03 | 1.77E-03 | 2.57E-03 | 1.67E-03 | 2.05E-03 | 2.62E-03 | 3.71E-03 | 1.83E-03 | 2.68E-02 | 3.01E-03 | |
| Patient-morphed | 3.84E-03 | 2.38E-03 | 1.56E-03 | 1.54E-03 | 2.45E-03 | 1.55E-03 | 1.96E-03 | 2.34E-03 | 3.39E-03 | 1.76E-03 | 2.56E-02 | 2.69E-03 | |
| Residual soft tissue | Patient-specific | 3.47E-03 | 3.37E-03 | 2.95E-03 | 2.38E-03 | 3.62E-03 | 2.61E-03 | 2.42E-03 | 2.68E-03 | 7.64E-04 | 3.17E-03 | 2.57E-03 | 6.15E-03 |
| Reference | 3.97E-03 | 3.79E-03 | 3.46E-03 | 2.78E-03 | 3.72E-03 | 3.14E-03 | 2.79E-03 | 3.33E-03 | 1.24E-03 | 3.61E-03 | 2.95E-03 | 6.70E-03 | |
| Patient-morphed | 3.82E-03 | 3.31E-03 | 3.02E-03 | 2.44E-03 | 3.32E-03 | 2.85E-03 | 2.49E-03 | 2.91E-03 | 1.06E-03 | 3.34E-03 | 2.63E-03 | 5.87E-03 | |
Figure 4:
Distribution of relative error in 18F S-value estimation. A) Relative difference in 18F S-values for retained organs of the non-deformed (i.e. reference) PSRK-Man/THRK-Man phantom, relative to S-values computed for the patient-specific segmentation-derived phantom. Values deviating by greater than 100% are shown in black. B) Relative differences for 18F S-values for the corresponding deformed phantom (relative to patient specific phantom). C) Box plot showing summary differences for S-values of the reference and deformed phantoms; whiskers represent the minimum and maximum relative differences. RST: residual soft tissue.
DISCUSSION
Routine nuclear medicine dosimetry is principally handled using reference phantoms, which assume the internal anatomy of specific patients is invariant and defined by population averages. The attenuation coefficient-based phantom deformation technique described herein was shown to improve the quality of anatomical representation of a reference phantom to better characterize a unique patient, based on multiple quantitative metrics of segment agreement. These improvements included enhanced spatial overlap accuracy (as-evaluated by overall-increased Dice similarity coefficients), improved agreement of organ boundary surfaces (as-evaluated by overall-reduced Hausdorff distances), and improved organ positioning (as-evaluated by reduced organ centroid positioning error), for organs of a deformed version of computational reference phantom relative to the unmodified (i.e. undeformed) control – each approach independently compared with a manual organ segmentation taken as ground truth.
In the present case, a number of structures from the reference and patient-morphed phantoms were merged with their respective sub-structures to enable more direct comparison with the patient-specific analog (e.g. walls/contents of heart, urinary bladder, and GI tract organs) or excluded from comparison due to inability to reliably segment manually based on CT contrast or resolution limitations (e.g. thyroid, lens of eye). However, it is critical to note that the present deformation methodology may be applied directly to a reference phantom in its native form, allowing all natively defined source/target regions to be retained – including (sub)regions too numerous to segment manually in practical clinical workflow, or too finely detailed (e.g. lens of eye, radiosensitive epithelia of the GI/respiratory tract) to define directly given limitations on scanner resolution. Figure 5 shows the reference and morphed phantoms utilized in the present study with all structures intact. Computational processing time was minimal (<5 minutes on a standard desktop PC) for all steps in the phantom deformation workflow, suggesting a bright outlook for the possibility of full automation in a clinical setting. Comparatively, manual segmentation required to generate the patient-specific phantom required approximately 2 hours. To note, for use in a practical dosimetry workflow, additional time (on the order of hours-to-days of cpu time)(7) would be required for Monte Carlo simulation using the derived phantom; however, this time would require no personnel interaction and may be drastically reduced with hardware acceleration/parallelization.
Figure 5:
Summary of advantages and limitations of phantom morphometric categories examined. Fully-detailed PSRK-/THRK-Man and deformed analog shown.
The set of S-values we generate here (Table 1) should be considered for comparative value. Despite the improved agreement of the deformed phantom based on these metrics and visual assessment, the algorithm as-utilized here is limited in its capability to improve dissimilarities between certain patient organ and reference phantom organ volumes – most notably for organs with small differences in linear attenuation coefficient relative to surrounding tissues, as well as for relatively small organs (e.g. gallbladder, pancreas) which may be stretched or compressed significantly upon large or sharp deformations of adjacent tissue. On the other hand, the algorithm performed well in the test case for organs with high CT contrast (e.g. bone, lungs, heart, body surface) and in general for larger organs (e.g. liver) whose overall volumes are less sensitive to change by deformation. Because organ self S-values depend heavily on organ mass, as is the case with, e.g., a reference phantom whose organs may differ significantly from a patient’s, we stress that it remains essential to carefully assess the implications of volume over- or underestimation when utilizing patient-morphed analogs in internal radionuclide exposures/radiopharmaceutical use cases. This will be of greater concern for isotopes with comparatively large self-dose contributions, i.e. therapeutic isotopes.
We note that in addition to sex- and age-matched reference phantoms, the present methodology may also be applied to height- and weight-percentile specific phantoms (i.e. patient-dependent phantoms). Selecting an initial phantom that already closely reflects patient body habitus and positioning would reduce computational overhead as well as the magnitude of potential error due to reduced deformation required. Comprehensive libraries of such phantoms are currently available, including a recently developed percentile-specific phantom series (20 phantoms)(18) and the UF/NCI phantom series (>400 phantoms)(19).
The specific deformable registration method utilized should be considered an implementation detail. We elected to use Plastimatch as the toolkit for deformable registration as it is implemented in 3D Slicer as a user-friendly plugin (alongside functionality for application of the output B-spline transforms to mesh geometry), and separately as a command-line tool, facilitating use of our approach by researchers both with and without programming expertise. We note that other deformable registration methods are certainly applicable, but the specific algorithm chosen should ideally be designed for co-registration of CT scans.
Several approaches to further improving the current methodology are under investigation, including utilization of additional imaging contrasts (e.g. magnetic resonance imaging, PET, or contrast-enhanced CT), and utilization of AI-assisted segmentation in combination with deformation.
CONCLUSION
In the present work, we sought to investigate the feasibility of an automatable approach to deforming pre-existing computational reference phantoms to better match patient anatomy, utilizing contrast similarity between patient CT scans and a reference phantom resampled based on reference values for organ attenuation coefficients. The test case evaluated demonstrates this technique can provide improvement in anatomical accuracy based on quality comparative metrics of organ overlap and surface distance, and provides proof-of-concept for developing patient-morphed phantoms from pre-existing reference phantoms, which could eventually be applied in clinical dosimetry workflows to attain a heightened degree of personalization, and by extension, greater accuracy in radiation dose estimation, without requiring time-consuming manual organ segmentation.
ACKNOWLEDGEMENTS AND DISCLOSURES
This research was funded in part through the NIH/NCI Cancer Center Support Grant P30 CA008748 and NIH U01 EB028234. We also acknowledge the Memorial Sloan Kettering Radiochemistry and Molecular Imaging Probes core, which was supported in part by NIH grant P30 CA08748. JSL acknowledges NIH R35 CA232130, the Mr. William H. and Mrs. Alice Goodwin and the Commonwealth Foundation for Cancer Research and The Center for Experimental Therapeutics of Memorial Sloan Kettering Cancer Center. LMC acknowledges support from the Ruth L. Kirschstein NRSA Postdoctoral Fellowship (NIH F32 EB025050). No other potential conflict of interest relevant to this article was reported.
REFERENCES
- 1.Mattsson S, Johansson L, Leide Svegborn S, et al. Radiation Dose to Patients from Radiopharmaceuticals: a Compendium of Current Information Related to Frequently Used Substances. Ann ICRP. 2015;44:7–321. [DOI] [PubMed] [Google Scholar]
- 2.Zvereva A, Schlattl H, Zankl M, et al. Feasibility of reducing differences in estimated doses in nuclear medicine between a patient-specific and a reference phantom. Phys Med. 2017;39:100–112. [DOI] [PubMed] [Google Scholar]
- 3.Clark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging. 2013;26:1045–1057. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Si H TetGen, a Delaunay-Based Quality Tetrahedral Mesh Generator. ACM Trans Math Softw. 2015;41:11:1–11:36. [Google Scholar]
- 5.Yeom YS, Jeong JH, Han MC, Kim CH. Tetrahedral-mesh-based computational human phantom for fast Monte Carlo dose calculations. Phys Med Biol. 2014;59:3173–3185. [DOI] [PubMed] [Google Scholar]
- 6.Sato T, Iwamoto Y, Hashimoto S, et al. Features of Particle and Heavy Ion Transport code System (PHITS) version 3.02. Journal of Nuclear Science and Technology. 2018;55:684–690. [Google Scholar]
- 7.Carter LM, Crawford TM, Sato T, et al. PARaDIM – A PHITS-based Monte Carlo tool for internal dosimetry with tetrahedral mesh computational phantoms. J Nucl Med. June 2019:jnumed.119.229013. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Patrick S, Birur NP, Gurushanth K, Raghavan AS, Gurudath S. Comparison of gray values of cone-beam computed tomography with hounsfield units of multislice computed tomography: An in vitro study. Indian J Dent Res. 2017;28:66–70. [DOI] [PubMed] [Google Scholar]
- 9.Kazerooni EA, Gross BH. Cardiopulmonary Imaging. Lippincott Williams & Wilkins; 2004. [Google Scholar]
- 10.Lepor H Prostatic Diseases. W.B. Saunders Company; 2000. [Google Scholar]
- 11.Paul Segars W, Tsui BMW. MCAT to XCAT: The Evolution of 4-D Computerized Phantoms for Imaging Research. Proc IEEE Inst Electr Electron Eng. 2009;97:1954–1968. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Zaffino P, Raudaschl P, Fritscher K, Sharp GC, Spadea MF. Technical Note: plastimatch mabs, an open source tool for automatic image segmentation. Medical Physics. 2016;43:5155–5160. [DOI] [PubMed] [Google Scholar]
- 13.Fedorov A, Beichel R, Kalpathy-Cramer J, et al. 3D Slicer as an Image Computing Platform for the Quantitative Imaging Network. Magn Reson Imaging. 2012;30:1323–1341. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med Imaging. 2015;15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Poulin E, Boudam K, Pinter C, et al. Validation of MRI to TRUS registration for high-dose-rate prostate brachytherapy. Brachytherapy. 2018;17:283–290. [DOI] [PubMed] [Google Scholar]
- 16.Zou KH, Warfield SK, Bharatha A, et al. Statistical Validation of Image Segmentation Quality Based on a Spatial Overlap Index. Acad Radiol. 2004;11:178–189. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Bolch WE, Eckerman KF, Sgouros G, Thomas SR. MIRD pamphlet No. 21: a generalized schema for radiopharmaceutical dosimetry--standardization of nomenclature. J Nucl Med. 2009;50:477–484. [DOI] [PubMed] [Google Scholar]
- 18.Lee H, Yeom YS, Nguyen TT, et al. Percentile-specific computational phantoms constructed from ICRP mesh-type reference computational phantoms (MRCPs). Phys Med Biol. January 2019. [DOI] [PubMed] [Google Scholar]
- 19.Geyer AM, O’Reilly S, Lee C, Long DJ, Bolch WE. The UF/NCI family of hybrid computational phantoms representing the current US population of male and female children, adolescents, and adults—application to CT dosimetry. Phys Med Biol. 2014;59:5225–5242. [DOI] [PMC free article] [PubMed] [Google Scholar]