Development of Scalable Lymphatic System in the 4D XCAT Phantom: Application to Quantitative Evaluation of Lymphoma PET Segmentations

Abstract

Background:

Digital anthropomorphic phantoms, such as the 4D extended cardiac-torso (XCAT) phantom, are actively used to develop, optimize, and evaluate a variety of imaging applications, allowing for realistic patient modelling and knowledge of ground truth. The XCAT phantom defines the activity and attenuation for a simulated patient, which includes a complete set of organs, muscle, bone, and soft tissue, while also accounting for cardiac and respiratory motion. However, the XCAT phantom does not currently include the lymphatic system, critical for evaluating medical imaging tasks such as sentinel node detection, node density measurement, and radiation dosimetry.

Purpose:

In this study, we aimed to develop a scalable lymphatic system in the XCAT phantom, to facilitate improved research of the lymphatic system in medical imaging. Using this scalable lymphatic system, we modelled the lymph node conglomerate pathology that is characteristically observed in primary mediastinal B-cell lymphoma (PMBCL). As an extended application, we evaluated PET image quantification of metabolic tumour volume (MTV) and total lesion glycolysis (TLG) of these simulated lymphomas, though the phantoms may be applied to other imaging modalities and study design paradigms (e.g., image quality, detection).

Methods:

A template model for the lymphatic system was developed based on anatomical data from the Visible Human Project of the National Library of Medicine. The segmented nodes and vessels were fit with non-uniform rational basis spline (NURBS) surfaces, and multichannel large deformation diffeomorphic metric mapping (MC-LDDMM) was used to propagate the template to different XCAT anatomies. To model conglomerates observed in PMBCL, lymph nodes were enlarged, converged within the mediastinum, and tracer concentration was increased. We used the phantoms as inputs to a PET simulation tool which generated images using ordered subsets expectation maximization (OSEM) reconstruction with 2–8 mm Gaussian filters. Fixed thresholding (FT) and gradient segmentation were used to determine MTV and TLG. Percent bias (%Bias) and coefficient of variation (COV) were computed as measures of accuracy and precision, respectively, for each MTV and TLG measurement.

Results:

Using the methodology described above, we introduced a scalable lymphatic system in the XCAT phantom, which allows for the radioactivity and attenuation ground truth to be generated in 116±2.5 seconds using a 2.3 GHz processor. Within the Rhinoceros interface, lymph node anatomy and function were modified to create a cohort of ten phantoms with lymph node conglomerates. Using the lymphoma phantoms to evaluate PET quantification of MTV, mean %Bias values were −9.3%, −41.3%, and 20.9%, while COV values were 4.08%, 7.6%, and 3.4% using 25% FT, 40% FT, and gradient segmentations, respectively. Comparatively for TLG, mean %Bias values were −27.4%, −45.8% and −16.0%, while COV values were 1.9%, 5.7% and 1.4%, for the 25% FT, 40% FT, and gradient segmentations, respectively.

Conclusions:

In this work, we upgraded the XCAT phantom to include a lymphatic system, comprised of a network of 276 scalable lymph nodes and corresponding vessels. As an application, we created a cohort of phantoms with lymph node conglomerates to evaluate lymphoma quantification in PET imaging, which highlights an important application of this work.

1. INTRODUCTION

Rapid developments in medical imaging have resulted in a wide variety of data acquisition and image generation methods available within the clinical setting^1–3. The implementation of these methods continues to exceed the pace by which these techniques can be properly validated and optimized. Clinical trials are slow to implement and often suffer from limited sample sizes, resources and funding⁴. Delineating the ground truth for human subjects, such as the location or volume of a lesion, is fundamentally a time-consuming and challenging task to perform within research studies^4,5. Phantoms are routinely used in imaging research and clinical practice to assess image quality and quantitative accuracy of images^6–10. However, few phantoms emulate the anatomical structures and heterogeneity features that are present in human subjects with a high degree of realism. Additionally, many phantoms do not provide users with the flexibility to modify anatomical structures^11–13; thus making it difficult to represent variation between patients. This may limit the ability to account for biological variability within the patient population and reduce generalizability towards clinical applications.

Virtual Clinical Trials (VCTs) provide a time-efficient solution for investigating research questions at a population-scale^14–17. In VCTs, the human subject is replaced with a digital phantom, imaging devices are simulated, and in certain cases, human observers are replaced with artificial observers. Like physical phantoms, digital phantoms have a known ground truth and can be used to evaluate and compare imaging devices and methods. Digital phantoms have the added opportunity to model human anatomy and physiology with an additional degree of sophistication. Voxelized phantoms¹⁸, such as the VIP-MAN^19,20, define anatomical structures based on prior image segmentations of patients. However, this approach results in rough organ boundaries and limits the representation of patients with different anatomy or pathology. Boundary representation (BREP) phantoms¹⁸ use advanced surface representations to model anatomical structures with higher degreees of realism. BREP uses mathematically defined surfaces to fit templates from patient segmentations. This results in realistic anatomical regions that can be easily manipulated to modify organ structure or account for temporal changes, such as patient motion due to respiration.

The 4D extended cardiac-torso (XCAT) phantom^21,22, a BREP-type phantom, allows for highly sophisticated multimodality imaging research. The XCAT phantom defines dozens of organs, accounts for tissue-type (e.g., muscle, bone, or soft-tissue), and allows users to modify demographic features such as age, sex, and body mass index (BMI), providing the opportunity for phantom studies to be performed at a population scale. Since the phantom is simulated, the attenuation and radioactivity ground-truth can be defined at a sub-voxel level, allowing for more sophisticated imaging metrics (e.g., shape, texture) to be evaluated. However, the XCAT phantom does not currently include the lymphatic system, relevant to tasks such as sentinel node detection^23,24, node density measurement^25,26, and radiation dosimetry^27,28. In this study, we aim to develop a scalable lymphatic system in the XCAT phantom, to facilitate improved research studies of the lymphatic system in medical imaging.

As an application, we considered primary mediastinal B-cell lymphoma (PMBCL), a potentially curable form of non-Hodgkin’s lymphoma, which classically presents with lymphadenopathy and bulky lymph node conglomerates located in the mediastinum^29,30. The International Prognostic Index (IPI) is used to predict outcome for non-Hodgkin’s lymphoma³¹. However, it is challenging to stratify risk for PMBCL patients with the IPI, as most patients are diagnosed at a young age and with disease restricted to the mediastinum^31,32. ¹⁸F-fluorodeoxyglucose ([¹⁸F]FDG) positron emission tomography/computed tomography (PET/CT) scans are increasingly used for initial assessment and monitoring of treatment response for lymphomas, including PMBCL³³. Baseline PMBCL tumour burden is often determined via [¹⁸F]FDG PET/CT, while post-treatment scans are used to determine tumour response to therapy³⁴, although there is evidence that quantitative imaging metrics can enhance the prognostic value of PET and its ability to effectively guide treatment decisions. For instance, metabolic tumour volume (MTV) and total lesion glycolysis (TLG) are well-documented predictors of therapy response and overall survival of lymphoma patients^35–39. Notably, a study by Ceriani et al. evaluated baseline PET/CT images of a PMBCL patient cohort stratified by TLG, and found that 5-year progression-free survival (PFS) for low and high TLG was 99% and 64%, respectively³³. However, the authors indicated that implementation of TLG in clinical practice would be premature, due to the lack of standardized guidelines and segmentation methods in the literature. Thus, there is substantial motivation to perform a quantitative evaluation of PET metrics within the context of lymphoma, such that they can be used to inform patient care.

Clinical reporting of MTV and TLG conventionally involves manual segmentation of lesions by a trained nuclear medicine physician. This approach is considered to be the gold-standard³⁶ but is time-consuming and often suffers from high intra- and inter-observer variability⁴⁰. Due to the need for standardized prediction tools, there is motivation to implement automated or semi-automated segmentation methods for routine clinical reporting of lymphoma. Two notable semi-automatic delineation methods include fixed thresholding (FT) and gradient-based segmentation. In PET imaging, FT selects voxels within a region which have radioactivity concentration equal to or greater than a certain threshold. This method is efficient to implement and largely observer-independent, resulting in low inter-observer bias³⁶. Due to the random nature of emissions from positron sources in the patient, FT algorithms are sensitive to noise and may be less robust⁴¹. Meanwhile, gradient segmentation methods utilize a contouring algorithm to define the tumour boundary^42,43. Although gradient segmentation algorithms have shown substantial advancement in recent years, these methods are highly vendor-specific and need to be carefully validated prior to introduction into a clinical setting³⁶. PET image quantification is also influenced by the reconstruction algorithm. Ordered subsets expectation maximization (OSEM) is conventionally used to reconstruct images, but a trade-off between image noise and image convergence exists depending on the number of iterations. Post-reconstruction filters are often used to “smooth” the image, but this comes at the expense of reduced image contrast and diagnostic sensitivity⁴⁴. As a result, there is motivation to efficiently evaluate these image processing and segmentation techniques prior to introduction in a clinical setting.

Overall, in this study, we describe the development of a scalable lymphatic system in the 4D XCAT phantom. Using this lymphatic system, we generate XCAT phantoms with lymph node conglomerates, to represent the pathology frequently observed in primary mediastinal B-cell lymphoma. As an application to quantitative imaging, we use the lymphoma phantoms as inputs to a PET simulation and reconstruction pipeline. Tumour uptake and volume will be evaluated using OSEM reconstruction with different Gaussian filter sizes and multiple noise realizations, to characterize the robustness and accuracy of fixed thresholding and gradient-based segmentation methods. Lastly, we consider tumour uptake and volume in patient images, to validate the generalizability of our phantom study in a clinical context.

2. MATERIALS AND METHODS

2.A. Development of the Lymphatic System in the XCAT Phantom

A template model for the lymphatic system was developed based on the anatomical data from the Visible Human Project of the National Library of Medicine (NLM)^45–47 as well as known anatomy. The lymphatic system dataset from the NLM consisted of 1878 transaxial anatomical slices with a matrix size of 2048×1216 with an in-plane pixel size of 0.33 mm and a slice thickness of 1 mm (Figure 1a). To define an initial model for the lymphatic vessels and nodes, the Visible Male anatomical data from the NLM were segmented using a combination of manual, automatic, and semi-automatic segmentation functions in the IMAGESEGMENT software program²¹ (Figure 1b). Polygon models of the segmented structures (Figure 2a) were output from IMAGESEGMENT and input to the Rhinoceros software package (www.rhino3d.com). The segmented vessels and nodes were fit with non-uniform rational basis spline (NURBS) surfaces using the surface lofting function within Rhinoceros: the software application used to create the original XCAT phantom (Figure 2b). Separate surfaces were created for each vessel and lymph node. The initial lymphatic system was applied based on known anatomy with the guidance of an expert physician. Additional vessels and nodes were added within Rhinoceros to define a total of 276 lymph nodes and corresponding vessels with NURBS surfaces in the XCAT phantom (Figure 2–c,d).

Figure 1:

Sample data used to define lymphatic system. (a) Example CT slice from the Visible Human Project of the National Library of Medicine (NLM). (b) Lymph nodes and vessel segmented from NLM dataset. Expanded image with annotated lymph nodes and vessels shown on the right.

Figure 2:

Development of lymphatic system template. (a) Polygon model for lymph node obtained from segmentation. (b) NURBS surface fit to polygon model. (c) Lymphatic system template defined using NURBS surfaces and supplemented based on known anatomy. (d) Extended cardiac-torso phantom (XCAT), in which the lymphatic system template is defined.

Given the template model, the multichannel large deformation diffeomorphic metric mapping (MC-LDDMM) method⁴⁸ was used to propagate the template to different XCAT anatomies. For this, the template model was voxelized into a 3D image with the organs and structures set to unique integer identifiers. Each target XCAT was then similarly voxelized with the same identifiers. Given the corresponding images, the MC-LDDMM method was used to calculate the transform from the template to the target. The image-based transform was then applied to the surface representation of the template lymphatic system to define it within the target. Transform vectors were simply applied to the control points defining each surface. Given the template and target images, MC-LDDMM calculates the transform between them; this transform is then applied to the model of the lymphatic system to define it in the given XCAT phantom. The MC-LDDMM method allows surfaces to be scaled according to patient gender, size, weight, age, and other anatomical differences, as shown in further detail in previous work by Segars et al.⁴⁹. For spatial overlap, the XCAT phantom has a priority system. Structures of higher priority will overwrite lower ones. Added objects such as the lymphatic system are set to the highest priority, so they overwrite anything in the background.

2.B. Application to Lymphoma Tumour Modelling

A pipeline was built to simulate human subjects with lymph node conglomerates in the mediastinum, as frequently observed in patients with PMBCL. A flow diagram illustrating the complete pipeline is shown in Figure 3, indicating the software and data format used at each step. Ten subjects were modelled using the XCAT phantom scripts⁵⁰ with our integrated lymphatic system: each subject had the same organ anatomy and uptake (height=176.1 cm, weight=81 kg), but exhibited variation in the lymph nodes. To model lymph node conglomerates, morphology and function were altered: lymph nodes were enlarged asymmetrically, converged within the mediastinum to form conglomerates, and tracer concentration was increased. This was achieved using commands in Rhinoceros to scale the x, y, and z dimensions of the lymph nodes, and translate them within the phantom. One lymph node conglomerate (consisting of 2–4 lymph nodes) was generated for each simulated patient, with tumour volumes ranging from 4 mL to 100 mL.

Figure 3:

Flowchart of lymphatic system development, tumour modelling, and image processing methodology.

2.C. PET Simulation and Image Processing

The XCAT general parameter script was used to input organ concentrations and generate binary files with ground-truth uptake and attenuation information. Input parameters for the radioactivity concentration of relevant organs are shown in Table 1. The mean radioactivity concentration for lymph node conglomerates (21.5 kBq/mL) was selected to correspond to the 50^th percentile of manually segmented lesions from a PMBCL patient analysis (22 lesions from 13 patients). As shown by ground truth images generated for the XCAT phantom, increased activity in the mediastinum can be observed for phantoms with lymph node conglomerates, as compared to non-pathological lymph nodes (Figure 4b and Figure 4c, respectively).

Table 1:

Input parameters used to define radioactivity concentration of organs in the XCAT phantom.

Region-of-Interest	Concentration [kBq/mL]
Background	1.5
Bladder	38.5
Esophagus (outer, contents)	4.1, 3.6
Heart (bloodpool, myocardium)	11.6, 55.0
Intestines	6.7
Kidney (medulla, cortex, pelvis)	19.8, 8.4, 25.3
Liver	13.1
Lung	3.5
Spleen	7.0

Figure 4:

Model of PMBCL disease pathology. (a) Lymph nodes expanded and converged within Rhinoceros viewing software, with additional organs shown. Coronal slice of radioactivity and attenuation distribution generated for patient with (b) primary mediastinal B-cell lymphoma and (c) non-pathological lymphatic system.

Ground-truth images of radioactivity and attenuation in the XCAT phantom (512×512 matrix size) were used as the input to a MATLAB-based PET simulation and reconstruction tool that generates simulated PET images for a GE Discovery RX scanner⁵¹ (Figure 5). Ten Poisson noise-realizations were generated for each subject. Images were reconstructed with ordered subsets expectation maximization (OSEM) using 2 iterations, 24 subsets, and 256×256 matrix size. Gaussian post-smoothing with different kernel sizes (2 mm, 4 mm, 6 mm, 8 mm) were applied and compared to images without post-smoothing (referred to as “Native” images). The files were converted from binary to DICOM file format and subsequently viewed using MIM (MIM Software, Inc.), a clinical radiology software program. The full image simulation pipeline and example files have been made publicly available (see Data and Code Availability at the end of manuscript).

Figure 5:

Coronal views using MIM. Bottom images include gradient-based PET Edge+ segmentations of tumour. (Left) [¹⁸F]FDG PET images of PMBCL patient with bulky mediastinal tumour. (Right) Simulated PET image using XCAT phantom. Tumour consists of 5 expanded lymph nodes with heterogeneous activity concentrations.

2.D. Segmentation Analysis

Regions-of-interest (ROIs) were delineated using MIM and saved in the RTStruct file format. DICOM images and the RTStructs were opened in Python and the rt-utils package (https://github.com/qurit/rt-utils) was used to convert the RTStructs into binary masks. Fixed thresholding (FT) (20%, 25%, 30%, 40%, 50%) segmentations were computed using our Python code. FT segmentation selects voxels with concentration greater than a certain threshold (e.g. % of SUV_max). Lesions were also segmented using MIM’s gradient-based algorithm (PET Edge+). The gradient method delineates tumour edges by calculating spatial derivatives along concentration line profiles^42,43. Both the FT and gradient-based tools provide time-efficient and reproducible methods for tumour segmentation. Metabolic tumour volume (MTV) and total lesion glycolysis (TLG) were calculated using each segmentation method.

To measure accuracy, percent bias (%Bias) for a given metric applied to a method (e.g., for MTV or TLG applied to tumors segmented using a particular method) was defined as:

$$\mathrm{\%}Bias=\frac{\frac{\sum _i^n{A}_{measured,i}}{n}-A_{truth}}{A_{truth}}\times 100\mathrm{\%}$$

where A_measured,i and A_truth are the measured and true metric values, respectively, of the ith noise realization. A_measured,i is averaged over n noise realizations (n = 10 in this study). Percent noise (coefficient of variation; COV), as a measure of precision, was in turn defined as:

$$COV=\frac{{\sigma }_{method}^A}{A_{truth}}\times 100\mathrm{\%}$$

where ${\sigma }_{method}^A$ is the standard deviation of the n measured values (e.g., MTV or TLG) for a given subject, as obtained from multiple noise realizations of the simulated patient.

2.E. Patient Study

Retrospective analysis of pre-treatment PET/CT scans was performed for 14 PMBCL patients. The injected activity for each patient ranged from 281–454 MBq and PET/CT scans were performed 60min post-injection. All scans occurred between October 2005 and February 2017. Fifteen mediastinal tumours were detected by a nuclear medicine physician and delineated using 25% FT, 41% FT, gradient (PET Edge+), and manual segmentation. %Bias and COV for MTV and TLG were computed and compared using manual segmentation as the ground truth. Bland Altman analysis was performed, and paired t-tests were used to compare MTV and TLG values for each segmentation method.

3. RESULTS

3.A. Development of XCAT Phantom for Lymphoma

Using the methodology described above, we introduced a scalable lymphatic system in the XCAT phantom. This updated phantom contains the same organs and structures present in the previous generation of the XCAT, but now includes a whole-body network of lymph nodes and vessels, as shown in Figure 2c. Using the Rhinoceros interface, the lymphatic system structures can be deformed and re-positioned by the user. As an application to lymphoma, a cohort of ten XCAT phantoms with lymph node conglomerates were created. Figure 6 shows a collage of the lymphatic system and relevant organs for each of these phantoms. Using the XCAT general parameter script, binary format files for the radioactivity and attenuation coefficients (512×512×301) were generated in 116.6±2.5 seconds with a 2.3 GHz processor.

Figure 6:

Coronal view of XCAT phantoms with lymph node conglomerates. (a.) Network of lymph nodes and vessels for each phantom. (b.) Rendering of lymphatic system with relevant organs in the mediastinum and abdomen.

While converting the NURBS surfaces to voxelized matrices, overlap between lymph nodes and adjacent structures (e.g., lung) becomes a relevant concern for lymph node conglomerates (Figure 6b). To ensure that all lymph nodes were included in the ground truth image, the lymphatic system was set to the highest priority, so they overwrite overlapping structures in the background. The structure priority list is included with the XCAT phantom files, such that these priorities may be amended by future users according to the specific research task and application. The updated XCAT phantom (i.e., with embedded lymphatic system) maintains the capability to be combined with previously developed SPECT/PET^51,52, CT^22,53, MRI^54,55, and ultrasound^56,57 simulation packages, many of which are freely available upon request.

3.B. Application to PET Segmentation

As an application to PET quantification of lymphoma, we used the newly developed XCAT phantoms as inputs to a PET simulation and reconstruction script. Figure 5 shows a simulated PET image of the XCAT phantom adjacent to the image of a patient scanned with [¹⁸F]FDG PET. For the phantom images, metabolic tumour volume (MTV) and total lesion glycolysis (TLG) values were compared using fixed thresholding (FT) and gradient-based segmentation applied to images filtered with a 6 mm post-smoothing kernel. Qualitatively, the gradient method smoothly followed the curvature of the lesion and sharp corners were not observed (Figure 7). By comparison, the FT method accepts only specific values of SUV_max based on the set percentage, and as such, jagged contours were not observed due to the voxel-by-voxel nature of the segmentation method. It should be noted that certain regions within the tumour appeared to be neglected for 40–50% FT.

Figure 7:

Simulated PMBCL tumour segmented with 20%, 25%, 30%, 40%, 50% fixed threshold (FT) and gradient-based PET Edge+ method.

MTV percent bias is plotted versus ground truth in Figure 8. MTV percent bias, obtained with 25% fixed thresholding, for the different tumour volumes was 1.3% (13 mL), 0.8% (39 mL), −13.1% (71 mL), and −11.9% (100 mL). Percent bias in MTV with PET Edge+ was 23.2% (13 mL), 14.1% (39 mL), 19.1% (71 mL), and 22.5% (100 mL). TLG percent bias versus ground truth is also plotted in Figure 8. The 25% thresholding led to percent bias of −20.8% at 13mL and −23.6% for the 100mL lesion. TLG percent bias using gradient segmentation was −15.2% at 13mL and deviated −11.8% for the 100mL lesion. As shown in noise (COV) versus bias plots (Figure 9), COV for MTV was 3.6%, 4.1%, 7.6%, and 3.4%, for 20% FT, 25% FT, 40% FT, and gradient segmentation, respectively. Similarly, COV for TLG was 1.4%, 1.9%, 5.7%, and 1.4%, for 20% FT, 25% FT, 40% FT, and gradient segmentation, respectively.

Figure 8:

Percent Bias+/−Percent Noise (COV) plotted versus ground truth for (top) MTV and (bottom) TLG metrics. Each colour corresponds to either a fixed threshold (FT) or gradient-based segmentation method.

Figure 9:

Percent noise (COV) vs. bias plots for (top) MTV and (bottom) TLG metrics. Each colour corresponds to either a fixed threshold (FT) or gradient-based segmentation method. Each color includes 10 points corresponding to ten different lymph node variations (ten subjects).

To investigate the effects of filtering on thresholding segmentation, images were smoothed with 2–8 mm Gaussian filters, and compared to the “Native” image (without smoothing). As shown in Figure S-1, 20–50% FT resulted in underestimation of MTV in the Native, 2 mm, and 4 mm smoothed images. However, MTV accuracy improved for images smoothed with 6 mm or 8 mm kernels, particularly if 20–30% FT are selected. TLG was consistently underestimated regardless of the selected threshold or kernel size. As shown in Figure S-2, utilizing larger kernel sizes (6 mm or 8 mm) reduced percent noise for 20–30% FT.

3.C. Validation with Patient Images

PMBCL tumours delineated using 25% FT, 41% FT, gradient, and manual segmentation are visually compared in Figure 10. As predicted by the simulated PMBCL images, the 41% FT appears to neglect certain regions of the tumours. The 25% FT, gradient, and manual segmentation show minor deviations from each other, but appear to be quite similar overall. Percent bias for MTV and TLG is plotted in Figure 11. Manual segmentation was used as the ground truth. Mean absolute error (MAE) for MTV was comparable for the 25% FT and gradient segmentation (127.7±34.1% and 116.7±40.1%, respectively). MAE was largest using the 41% FT (141.6±38.9%). MAE computed for TLG was similar for the 25% FT and gradient methods (118±35.6% and 115.5±34.8%), and largest for 41% FT (130.5±40.1%).

Figure 10:

PMBCL tumour delineated with 41% and 20% fixed threshold (FT), gradient, and manual segmentation. MTV indicated for each tumour (determined via manual segmentation).

Figure 11:

MTV and TLG percent bias for 25% FT, 41% FT, and gradient-based segmentation methods, using manual segmentation as the ground truth.

Percent difference between gradient and 25% FT segmentation methods is shown using Bland Altman plots (Figure S-3). Percent Bias±Percent Noise for MTV and TLG were 1.5±78.9% and −1.4±69.5%, with range −135.5–148.7% and −124.4–141.9%, respectively. Mean percent bias did not differ significantly from zero (P=0.944 and P=0.941). However, percent difference for larger lesions (MTV>100 mL or TLG>1000 kBq) significantly differed from zero (P=3.9 × 10⁻³ and P=0.012), and mean percent bias were 28.2±8.2% and 15.2±9.7% for MTV and TLG, respectively.

4. DISCUSSION

In this work, we upgraded the XCAT phantom to include a lymphatic system, comprised of a network of 276 deformable lymph nodes and corresponding vessels. Through controlled manipulation of lymphatic anatomy and function, lymph node pathology can be scaled and evaluated in digital experiments. Distributed to the research community, this upgraded XCAT phantom has potential for improved quantification and detectability studies in medical imaging. For instance, a scalable lymphatic system is critical for modelling diseases with abnormal lymph anatomy or function, such as lymphoedema^58,59, lymphangiogenesis⁶⁰, or lymphoma^61,62.

In this study, we used boundary representation methods to fit the lymph nodes and vessels with non-uniform rational basis spline surfaces (NURBS). The primary motivation for using NURBS surfaces (as opposed to voxelized structures) is that it provides a flexible framework to manipulate lymph nodes and vessels while accounting for temporal changes, such as patient motion due to respiration⁶³. This is particularly advantageous for virtual clinical trials which seek to define phantom populations with minimal computational time. Coupled with the ML-DDMM method, there remains the capability to transform the lymphatic system for patients with varying height, weight, and ages. Using this approach, each object can be re-positioned and scaled with three degrees of freedom. By selecting input values in the general parameter script, the radioactivity and attenuation for each lymph node and vessel can be independently defined (Figure 3). This allows flexibility while designing experiments to represent anatomical and pathological variations of the lymphatic system.

We do recognize some limitations within this study. Although anatomical deformations in the XCAT phantom are handled automatically, scaling and re-positioning of lymph nodes and vessels is performed manually using the Rhinoceros software. This increases the preparation time required to model patients with different lymphatic pathologies, which is a potential hurdle for large-scale virtual clinical trials. Future efforts will allow for lymph positioning and scaling to be dually controlled using the Rhinoceros software and XCAT general parameter script. It is important to note that we have developed the lymphatic system for the adult XCAT phantom, though we have yet to extend the structures to represent pediatric populations. Due to the substantial difference in nodal size and distribution between pediatric and adult populations⁶⁴, further work is required to define the lymphatic system using separate template images from the NLM.

In this study, it is also necessary to discuss similar developments by Lee et al.⁶⁵, whose work simulated lymph nodes in a virtual phantom in order to estimate parameters for radiation dosimetry. To achieve this, Lee et al. used a combination of NURBS and polygon-mesh surfaces to model lymph nodes in 16 different sites across the body, for digital phantoms designed to represent pediatric and adult patients. A random sampling method was used to position the lymph nodes in various sites of the body, until the target node mass was fulfilled. In contrast, we defined the lymphatic system based on the sample dataset from the NLM and known anatomy, which gives us the ability to individually design the size, shape, and location of each lymph node and vessel in the phantom. Additionally, our work introduced scalable lymph nodes and vessels, which is critical for modelling disease pathologies such as lymphoedema^58,59 and lymphangiogenesis⁶⁰, respectively.

As described above, the development of the lymphatic system allows us to model new pathology in the XCAT phantom, such as the bulky lymph node conglomerates in lymphoma. As [¹⁸F]FDG PET is an important staging tool for primary mediastinal B-cell lymphoma (PMBCL)^30,33,39, there is opportunity to incorporate quantitative imaging features, such as tumour volume and glycolysis, into routine clinical decision-making. However, rapid developments in medical imaging have resulted in a wide variety of data acquisition and image generation methods available within the clinical setting^1–3, urging the need for further standardization and validation. Thus, the upgraded XCAT phantom provides an opportunity to evaluate MTV and TLG quantification for PMBCL with precisely defined ground truth.

Within our study, as shown by visual comparison of the simulated segmented tumours (Figure 7), selecting a 20–30% FT appears to result in accurate delineation of the tumour. These results were validated quantitatively in Figure 8. MTV quantification using 25% FT was most accurate for lesions <50mL, while the 20% FT was best for lesions >50mL. All FT segmentations underestimated total lesion glycolysis. TLG error for 20–30% FTs was within 30 percent. The 40% and 50% thresholding typically had percent biases greater than 30 percent. Therefore, selecting thresholds between 20% and 30% appears to be most suitable for MTV and TLG quantitation. Our results found that 40% thresholding was suboptimal, due to internal regions in the tumour not being selected in the segmentation. These findings are in agreement with those made by Sridhar et al.⁴³ in head and neck cancer, in which 40% and 50% FT were shown to have poor correlation with pathological tumour volume (unlike gradient based segmentation). This contrasts with the European Association of Nuclear Medicine (EANM) recommendation, which suggests 41% FT for tumour segmentation⁶⁶. It is important to note that our results are possibly specific to the pathology of lymphoma, which characteristically exhibits heterogeneous tumour shapes and tracer uptake patterns. It should be noted that our simulation models the GE Discovery RX scanner. Care should be taken to validate these results with different image reconstruction and processing methods, and on a diverse range of scanners in the field.

Specifically for PMBCL, Meignan et al.⁶⁷ observed that 41% FT is most accurate for determining MTV. Our contrasting results could be related to the methodology used within the phantom experiment. The lymphatic systems of our digital phantoms were modified to reflect the pathological properties of non-Hodgkin’s lymphoma, which is known to exhibit tumours composed of lymph node conglomerates. Comparatively, Meignan et al.⁶⁷ modelled tumours using anatomy that has been previously observed in PET images, by injecting radioactivity into saline bags. This highlights an opportunity created by the upgraded XCAT phantom to model pathological properties of lymphoma, rather than observed anatomy.

Gradient-based segmentation methods have shown recent advancements in lesion segmentation tasks, although these methods are highly vendor-specific and need to be carefully validated prior to introduction into a clinical setting³⁶. On average, the specific gradient-based segmentation tool evaluated within our study delineated larger volumes than the 20% FT segmentation. In fact, the gradient-based method consistently overestimated MTV (Figure 8). By contrast, the gradient-based method was reasonably accurate for determining TLG. These were encouraging results, since previous studies have observed gradient-based methods to have poor performance for heterogeneous or low-uptake tumours³⁶, which is characteristic of PMBCL. The overestimated tumour boundaries for the gradient method partially compensated for spillover due to the partial volume effect⁶⁸. Given the development of partial volume correction (PVC) strategies in PET^69–71, it would be interesting for future studies to evaluate the effect of PVC on MTV and TLG quantification.

To investigate the generalizability of our phantom study findings, patient images of PMBCL patients were also evaluated for comparison purposes. As shown in Figure 11, the 25% FT and gradient-based segmentation methods determined similar MTV and TLG, as compared with the manual segmentation. The recommended 41% FT underestimated MTV and TLG, compared to the manual segmentation. Given these results, it appears that 41% FT may not be optimal for PMBCL segmentation tasks. Due to the high degree of variability associated with manual segmentation³⁶, it is important to consider that these findings may not generalize to different scanners and reconstructions. Further investigation is required to confirm whether this trend among segmentations is consistent for multiple independent observers. The relationship between 25% FT and gradient segmentation was most consistent for bulky mediastinal tumours (>100mL). The gradient method consistently estimated larger MTV and TLG values than the 25% FT. These findings were predicted from the simulated PET/CT images. However, the percent difference for smaller tumours (<100mL) did not follow a consistent trend. The inconsistent performance for smaller tumours may be related to biological variation between patients, which is notable in PMBCL.

Variation in tumour location, shape, and tracer uptake may uniquely impact each segmentation method. For instance, cold-spots in the tumour may uniquely influence the 25% FT, which segments on a voxel-scale, while the gradient method may not be impacted as it depends on larger-scale uptake patterns within the tumour^42,43. In agreement with the simulated experiment, gradient segmentation resulted in larger tumour volume and activity values, compared to the 25% FT. We found that total lesion glycolysis was more robust against Poisson noise than metabolic tumour volume; however, it is also necessary to consider the biological and pathological reproducibility of each metric prior to introduction in a clinical setting. Overall, our results suggest usage of 20–25% FT for delineating tumour volume, and gradient-based segmentation for determining total lesion glycolysis. Given our paradigm, future efforts can include assessment of accuracy and reproducibility for a range of higher-order imaging (radiomics) features.

5. CONCLUSIONS

In this study, we describe the development of a scalable lymphatic system in the 4D XCAT phantom, which was comprised of a network of 276 deformable lymph nodes and corresponding vessels. The lymphatic system was defined using non-uniform rational basis spline (NURBS) surfaces, allowing for each lymph node and vessel to be independently scaled and translated for different XCAT anatomies. As an application to quantitative evaluation of PET segmentations, we modified the lymphatic system to represent the bulky lymph node conglomerates that are frequently observed in primary mediastinal B-cell lymphoma. Distributed to the research community, the lymphatic system has the potential to enable image quality and quantification studies in medical imaging, particularly for disease pathologies that are relevant to the lymphatic system.

Supplementary Material

fS3

Figure S-3: Bland-Altman analysis plotting percent difference (Δ%) vs. average of gradient and 25% FT segmentation. TMTV and TLG are displayed for full dataset (top) and subset of data (bottom). 95% CI (dotted lines) and mean percent bias (solid lines) are shown.

fS2

Figure S-2: Percent noise (COV) vs. bias plots for (top) TMTV and (bottom) TLG metrics. Each column of plots corresponds to images with different post-smoothing kernel sizes applied. Each colour corresponds to 20%, 25%, 30%, 40%, 50% fixed thresholding (FT) segmentation. Each color includes 10 points corresponding to ten different lymph node variations (ten subjects).

fS1

Figure S-1: Percent Bias±Percent Noise (COV) is plotted versus ground truth for (top) TMTV and (bottom) TLG metrics. Each column corresponds to images with different post-smoothing kernel sizes. Each colour corresponds to 20%, 25%, 30%, 40%, 50% fixed thresholding (FT) segmentation.

DATA AND CODE AVAILABILITY

All codes (including PET simulation and reconstruction algorithms, dicom conversion scripts) and simulated datasets are shared publicly at: https://Qurit.ca/software and https://github.com/qurit/xcat-sims

The upgraded XCAT phantom is available upon request from Dr. Paul Segars: paul.segars@duke.edu and additional information can be accessed at: https://olv.duke.edu/industry-investors/available-technologies/xcat/