The Open Collections website will be unavailable July 27 from 2100-2200 PST ahead of planned usability and performance enhancements on July 28. More information here.

UBC Undergraduate Research

Quantitative PET/CT Imaging for Enhanced Evaluation of Prostate Cancer and Lymphoma Fedrigo, Roberto 2021-04

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata


52966-Fedrigo_Roberto_PHYS_449_PET_CT_imaging_2021.pdf [ 10.55MB ]
JSON: 52966-1.0400082.json
JSON-LD: 52966-1.0400082-ld.json
RDF/XML (Pretty): 52966-1.0400082-rdf.xml
RDF/JSON: 52966-1.0400082-rdf.json
Turtle: 52966-1.0400082-turtle.txt
N-Triples: 52966-1.0400082-rdf-ntriples.txt
Original Record: 52966-1.0400082-source.json
Full Text

Full Text

Quantitative PET/CT Imaging for Enhanced Evaluation ofProstate Cancer and LymphomabyRoberto FedrigoA THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFBachelor of Science in Honours BiophysicsinTHE FACULTY OF SCIENCE(Physics & Astronomy)The University of British Columbia(Vancouver)April 2021© Roberto Fedrigo, 2021AbstractPositron Emission Tomography (PET) scans offer valuable insight into the progres-sion of diseases, such as cancer. However, there is evidence that quantitative imag-ing metrics can enhance the prognostic value of PET and its ability to effectivelyguide treatment decisions. This thesis aims to improve the accuracy and repro-ducibility of tumour metrics as determined through PET images. To achieve this,we implemented both physical and simulated experiments using anthropomorphicphantoms. We evaluated two contrasting domains within nuclear medicine: (a.)Prostate cancer, in which prostate-specific membrane antigen (PSMA) PET imag-ing is used for detection of focal lesions, as well as (b.) Primary mediastinal B-celllymphoma, in which [18F]FDG PET is used to detect bulky, heterogeneous lymphnode conglomerates.For PSMA PET imaging, we propose and evaluate SUVapex, a segmentation-free metric for determining tumour tracer concentration. We also introduce theCanadian PET Prostate Phantom for Oncology (C3PO) - a PET/CT/MRI-compatiblephantom designed for harmonization of PSMA PET imaging. Within the context ofPMBCL PET, our results suggest that the 25% fixed threshold provides better ac-curacy for tumour volume quantification. Ideally, this study will lead to improvedclinical reporting of tumour volume, which will result in improved outcome pre-diction and disease management of lymphoma within British Columbia.iiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiGlossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Prostate Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Lymphoma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.3 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 Theory of Nuclear Medicine Imaging . . . . . . . . . . . . . . . . . 72.1 Positron Emission Tomography . . . . . . . . . . . . . . . . . . . 72.2 Phantom Applications in PET . . . . . . . . . . . . . . . . . . . 82.3 Reconstruction Algorithms . . . . . . . . . . . . . . . . . . . . . 112.4 Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . 122.5 Image Quantification . . . . . . . . . . . . . . . . . . . . . . . . 123 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153.1 Quantitative Evaluation of Prostate Cancer . . . . . . . . . . . . . 15iii3.1.1 PSMA Patient Analysis . . . . . . . . . . . . . . . . . . . 153.1.2 Shell-less 22Na Lesions . . . . . . . . . . . . . . . . . . . 153.1.3 Probe-IQ Phantom Image Acquisition . . . . . . . . . . . 173.1.4 Data Processing and Analysis . . . . . . . . . . . . . . . 173.1.5 Canadian PET Prostate Phantom for Oncology . . . . . . 183.2 Non-Hodgkin’s Lymphoma Quantification . . . . . . . . . . . . . 213.2.1 Physical Phantom Experiments . . . . . . . . . . . . . . . 213.2.2 Simulated Phantom Experiments . . . . . . . . . . . . . . 234 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264.1 PSMA Experiments . . . . . . . . . . . . . . . . . . . . . . . . . 264.1.1 Lesion Segmentations . . . . . . . . . . . . . . . . . . . 264.1.2 Recovery Curves . . . . . . . . . . . . . . . . . . . . . . 264.1.3 Robustness of Recovery Curves . . . . . . . . . . . . . . 284.2 Lymphoma Quantification . . . . . . . . . . . . . . . . . . . . . 304.2.1 Physical Phantom Experiments . . . . . . . . . . . . . . . 304.2.2 Simulated Phantom Experiments . . . . . . . . . . . . . . 305 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385.1 Applications to Prostate Cancer . . . . . . . . . . . . . . . . . . . 385.1.1 PSMA Lesion Quantification . . . . . . . . . . . . . . . . 385.1.2 PSMA PET Harmonization Study . . . . . . . . . . . . . 405.2 Applications to Lymphoma . . . . . . . . . . . . . . . . . . . . . 405.2.1 Lymphoma Simulation . . . . . . . . . . . . . . . . . . . 405.2.2 Physical Phantom Experiment . . . . . . . . . . . . . . . 415.3 Advancements in Phantom Technology . . . . . . . . . . . . . . 425.4 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . 445.4.1 Dosimetry for Radiopharmaceutical Therapy . . . . . . . 445.4.2 Outcome Prediction for Non-Hodgkin’s Lymphoma . . . . 456 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476.1 PSMA PET Quantification . . . . . . . . . . . . . . . . . . . . . 476.2 PMBCL Quantification . . . . . . . . . . . . . . . . . . . . . . . 47ivBibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49A Supporting Materials . . . . . . . . . . . . . . . . . . . . . . . . . . 54A.1 C3PO Full Study Protocol . . . . . . . . . . . . . . . . . . . . . 54vList of TablesTable 3.1 Measured [18F]FDG activity concentrations for PMBCL patientsin different regions of interest. . . . . . . . . . . . . . . . . . . 23viList of FiguresFigure 1.1 Conventionally, scanner performance is validated using NEMAImage Quality (IQ) phantom. [18F]FDG is injected into fillablespheres to simulate lesions, as highlighted by the blue dye. . . 3Figure 1.2 Primary mediastinal B-cell lymphoma (PMBCL) patient withbulky primary tumour located in the mediastinum. Lesion andliver segmentations shown in images on the right. Patient im-aged using [18F]FDG PET (top) and [18F]FDG PET/CT (bottom). 4Figure 1.3 Overview of research study pipeline to evaluate improve quan-titative metrics for PMBCL lymphoma. . . . . . . . . . . . . 6Figure 2.1 a.) Positron annihilation: Positron Emission Tomography(PET) tracers emit positrons, which travel through the sur-rounding medium until they annihilate with an electron. Two511 keV γ-rays are emitted from the annihilation location atapproximately 180◦ directions. Adapted from Chandra andRahmim [1]. b.) Detector ring: PET scanners contain de-tector rings which detect 511 keV γ-rays. If two γ-rays aredetected within a small time-frame then they are assumed tohave occurred from same annihilation event. Adapted fromUribe, 2019 [2]. . . . . . . . . . . . . . . . . . . . . . . . . . 9Figure 2.2 (a.) NEMA Image Quality (IQ) phantom. (b.) Probe-IQ phan-tom. (c.) Simulated 4D-extended cardiac torso (XCAT) phantom. 10Figure 2.3 (Left-to-right) Manual, fixed threshold (FT), and gradient-basedsegmentation methods. . . . . . . . . . . . . . . . . . . . . . 13viiFigure 3.1 Transaxial PET image slices of 22Na epoxy lesions in [18F]FDGbackground. Different lesion contrast is observed initially (t=0)compared to 4 hours later. . . . . . . . . . . . . . . . . . . . 16Figure 3.2 (a.) Schematic of aluminum mold used for casting 3-16mmspheres. (b.) Radioactive epoxy spheres (3-16mm) infusedwith [22Na]NaCl. . . . . . . . . . . . . . . . . . . . . . . . . 16Figure 3.3 (Top) Medium and large phantom shell. (Bottom) Pelvic phan-tom compartment with bladder insert, as well as ribs and spinethat can be inserted in the large phantom. . . . . . . . . . . . 18Figure 3.4 (a.) PET image of metastatic prostate cancer patient imagedwith [18F]DCFPyL, targeting Prostate-Specific Membrane Anti-gen. (b.) PET image of Probe-IQ phantom. . . . . . . . . . . 19Figure 3.5 Canadian PET Prostate Phantom for Oncology. (a.) Externalview of phantom. (b.) View of opened phantom to show epoxylesions. (c.) Acrylic cylinders that can be inserted into phantom. 20Figure 3.6 Probe-IQ lymphoma experiment pipeline. (a.) Segmentationof bulky PMBCL tumour saved in a stereolithography file for-mat. (b.) Radioactive PMBCL tumour model casted withliquid plastic. (c.) PET/CT coronal slice of PMBCL tumourmodel inserted into Probe-IQ thorax shell. . . . . . . . . . . . 22Figure 3.7 Lymphatic system development pipeline. (a.) Anterior view ofmale (left) and female (right) 4D XCAT phantom anatomies.Adapted from figure by Segars et al., 2008. (b.) Lymphaticsystem incorporated for XCAT phantom, which can be com-bined with XCAT organs. (c.) Coronal slice of combinedXCAT and lymphatic phantom showing radioactivity and at-tenuation distribution for a ”healthy” patient. . . . . . . . . . 24Figure 3.8 PMBCL tumour simulation. (Left) Lymph nodes expanded andconverged within Rhinoceros 3D viewing software, with addi-tional organs shown. (Right) Coronal slice of radioactivity andattenuation file generated for patient with bulky mediastinaltumour. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25viiiFigure 4.1 Segmentation methods applied to PET images of 14mm, 8mm,and 6mm 22Na epoxy spheres. . . . . . . . . . . . . . . . . . 27Figure 4.2 (Top to bottom) Recovery concentration coefficients measuredin Probe-IQ pelvis using Max, Peak, Apex, and Mean (40% FTand gradient). (Left to right) Reconstruction algorithms usingOSEM+PSF) and BSREM (32 subsets). Mean Absolute Error(MAE) +/- Standard Deviation indicated on each plot. . . . . 29Figure 4.3 Recovery concentration coefficient versus lesion-to-backgroundratio for 10mm lesion. (Top to bottom) Max, Peak, Apex, andMean (40% FT and gradient). (Left to right) Reconstruction al-gorithms using OSEM+PSF and BSREM (32 subsets). Rangeand standard deviation of recovery coefficients annotated onplots. Realistic PSMA background activity concentration de-noted by pink shaded region. . . . . . . . . . . . . . . . . . . 31Figure 4.4 PET image of PMBCL tumour model inserted into Probe-IQphantom. Tumour segmented using 20%, 25%, 30%, 40%,50% fixed threshold (FT) and gradient-based PET Edge+ method. 32Figure 4.5 Percent bias vs. true value plotted for: (Top) Total metabolictumour volume (TMTV) and (Bottom) Total lesion glycolysis(TLG) plotted versus true value, using PMBCL tumour modelscasted using [18F]FDG-infused liquid plastic, with 1 minutebed duration. . . . . . . . . . . . . . . . . . . . . . . . . . . 33Figure 4.6 Coronal views of PET images using MIM. Bottom images in-clude gradient-based segmentation of tumour. (Left) [18F]FDGPET image of PMBCL patient with bulky mediastinal tumour.(Right) Simulated PET image using XCAT phantom. Tumourconsisted of 5 expanded lymph nodes with heterogeneous ac-tivity concentrations. . . . . . . . . . . . . . . . . . . . . . . 34Figure 4.7 PET image of PMBCL tumour simulated using XCAT phan-tom. Tumour segmented using 20%, 25%, 30%, 40%, 50%fixed threshold (FT) and gradient-based PET Edge+ method. . 36ixFigure 4.8 Percent bias vs. true value plotted for: (Top) Total metabolictumour volume (TMTV) and (Bottom) Total lesion glycoly-sis (TLG) plotted versus true value. Lesions consisted of 2-4asymmetric lymph nodes converged within the mediastinum,and ranged from 4-100mL. . . . . . . . . . . . . . . . . . . . 37Figure 5.1 (a.) Conventional NEMA Image Quality phantom. (b.) Cana-dian PET Prostate Phantom for Oncology. (c.) Image qualityprobe (Probe-IQ) phantom. (d.) Simulated 4D-extended car-diac torso phantom (XCAT). (e.) Non-Hodgkin’s lymphomapatient imaged with PET. . . . . . . . . . . . . . . . . . . . . 44Figure A.1 Schematic of Canadian PET Prostate Phantom for Oncology.Phantom visualized in 3D. . . . . . . . . . . . . . . . . . . . 56Figure A.2 Schematic of Canadian PET Prostate Phantom for Oncology.Size and shape of each region is visualized. . . . . . . . . . . 57Figure A.3 Schematic of Canadian PET Prostate Phantom. Design of eachacrylic cylinder is visualized. . . . . . . . . . . . . . . . . . . 58xGlossaryBSREM Block sequential regularized expectation maximizationC3PO Canadian PET Prostate Phantom for OncologyFT Fixed threshold segmentationNCMO Negative-Cast Modelling for OncologyOSEM Ordered subset expectation maximizationPMBCL Primary mediastinal B-cell lymphomaXCAT 4D-extended cardiac torso phantomxiAcknowledgmentsI would like to thank everyone in the Quantiative Radiomolecular Imaging & Ther-apy Lab for their support. Furthermore, I would like to take the opportunity tothank my supervisors, Dr. Carlos Uribe and Dr. Arman Rahmim. This thesiswould not have been possible without their thoughtful mentorship and guidance.Thank you to Dr. Paul Segars, Dr. Patrick Martineau, Dr. Robin Coope, aswell as Dr. Dan Kadrmas, Dr. Francois Benard, Dr. Peter Petric, Lauren Fougner,and Dr. Patricia Edem for their collective efforts and collaboration towards theculmination of this thesis.Finally, I would like to thank my parents and my brother, for their constantsupport throughout my undergraduate education and life.This project was in part supported by Natural Sciences and Engineering Re-search Council (NSERC) of Canada Discovery Grant RGPIN-2019-06467 and theCanadian Institutes of Health Research (CIHR) Project Grant PJT-162216.xiiChapter 1IntroductionCancer is one of the leading causes of death worldwide [3]. It is estimated that 1 inevery 2 Canadians will be diagnosed with cancer in their lifetime, while 225,800Canadians were diagnosed in 2020 alone [4]. Cancer is defined by the NationalInstitute of Health as a disease in which ”abnormal cells divide without controland can invade nearby tissues” [5]. Essentially, cancer is a term used to classifyover 200 diseases with highly heterogeneous epidemiology [6]. Early detection ofcancer has been shown to significantly reduce the chances of disease progressionand death [7]. Medical imaging contains some of the most effective methods forearly detection and diagnosis of cancer. The various approaches to imaging can becategorized into three major groups:• Structural imaging: Identifies regions based on anatomical properties. Forinstance, computed tomography (CT) differentiates structure based on tis-sue density. Unfortunately, some malignancies do not exhibit an observablechange in structure, thus preventing detection.• Functional imaging: Provides detailed information regarding biologicalprocesses, such as metabolism. For example, functional magnetic resonanceimaging (fMRI) measures blood flow to determine processes related to brainactivity.• Molecular imaging: Molecular imaging is a subset of functional imagingthat allows for direct visualization of molecular processes in vivo. For ex-1ample, positron emission tomography (PET) uses radiopharmaceuticals tomonitor biochemical interactions. Images with PET are characteristicallynoisy and have low-resolution, but have high sensitivity.This thesis is focused on the use of Positron Emission Tomography (PET)within oncology. Specifically, we will consider PET imaging within the contextof prostate cancer and lymphoma.1.1 Prostate CancerProstate cancer (PC) is the second most prevalent malignancy in men and fifthdeadliest worldwide [8]. Although locally contained prostate cancer has high(>90%) survival rates, metastatic prostate cancer (mPC) exhibits extremely lowsurvival after 5 years [8]. Positron emission tomography (PET) with18F-fluorodeoxyglucose ([18F]FDG) is frequently used to detect mPC, although itis challenging to detect the sub-cm tumours that are characteristic of the disease.A new generation of PET tracers that target the prostate specific membrane anti-gen (PSMA) have shown improved specificity for binding to prostate cancer cells.For instance, PET images with the popular PSMA-targeting tracer, [18F]DCFPyL,have shown superior results in detecting prostate cancer [9, 10] and resulted in theobservation of high contrast and focal lesions.Accurate quantification of lesions imaged with PSMA PET can enable eval-uation of therapeutic efficacy, harmonization between sites, and the potential tobuild outcome predictive models. Conventionally, scanner performance is vali-dated using the NEMA Image Quality phantom (Fig. 1.1), in which [18F]FDG isinjected into 10-37mm fillable spheres, at 4:1 lesion-to-background activity ratios.The NEMA approach, while ‘easy’ and reproducible, does not approximate thehigh contrast, low-diameter lesions that are characteristic of PSMA-based imagingagents. Furthermore, the NEMA approach can create “cold shells” of activity be-cause of the thickness of the plastic walls of the spheres; not representative of reallesions. The relative volume of a “cold shell” becomes significant for small diame-ter lesions, and has been found to reduce the measured concentration recovery [11]and increase the observed volume of a lesion [12]. Alternate methods have been2Figure 1.1: Conventionally, scanner performance is validated using NEMAImage Quality (IQ) phantom. [18F]FDG is injected into fillable spheresto simulate lesions, as highlighted by the blue dye.developed to circumvent the cold shell effect [11, 13, 14], although few studieshave evaluated quantification for focal, high contrast lesions. Therefore, we aim tomodel tumours within this emerging context of PSMA PET imaging.1.2 LymphomaPrimary mediastinal B-cell lymphoma (PMBCL) represents a potentially curableform of Non-Hodgkin’s lymphoma, which classically presents with bulky, het-erogeneous tumour masses located within the mediastinum [15] (Fig. 1.2). Thegold-standard of treatment involves immunochemotherapies combined with tar-geted therapy [16]. Unfortunately, this cancer varies significantly within the popu-lation and each patient responds differently to treatment, making disease manage-ment and decision making particularly challenging.Positron Emission Tomography (PET) scans offer valuable insight into the pro-gression of PMBCL. Through the development of specially-designed radiophar-maceuticals, complex biological processes can be monitored in vivo. For instance,differential uptake of glucose can be visualized using 18F-fluorodeoxyglucose3Figure 1.2: Primary mediastinal B-cell lymphoma (PMBCL) patient withbulky primary tumour located in the mediastinum. Lesion and liversegmentations shown in images on the right. Patient imaged using[18F]FDG PET (top) and [18F]FDG PET/CT (bottom).4([18F]FDG), a glucose analog. As a result, [18F]FDG PET/CT imaging is fre-quently used to quantify glucose metabolism and provide valuable informationregarding disease prognosis.Baseline PMBCL tumour burden is often determined via [18F]FDG PET/CT,while post-treatment scans are used to determine tumour response to therapy [17].The Deauville 5-point scale (Deauville 5PS) is an internationally-recommendedstaging system for clinical reporting of PMBCL [18]. Tracer uptake of each [18F]FDG-avid tumour is assessed semi-quantitatively, relative to the mediastinum and liveruptake using a 1-to-5 scale. However, there is evidence that quantitative imagingmetrics can enhance the prognostic value of PET and its ability to effectively guidetreatment decisions.Total metabolic tumour volume (TMTV) and total lesion glycolysis (TLG) arewell-documented predictors of therapy response and overall survival of lymphomapatients [19–22]. Unfortunately, reporting of TMTV and TLG requires manualdelineation by trained nuclear medicine physicians. Because segmentation is avery time-consuming process, these metrics are not routinely reported. Thus, thereis significant motivation to implement artificial intelligence (AI) assisted meth-ods (e.g. deep learning) for automated segmentation in a clinical setting such thatreporting TMTV becomes a standard of care to better tailor therapies for eachindividual.Tumour quantification accuracy can be impacted by the selected reconstructionparameters and segmentation algorithms. Imaging phantoms (specially-designedobjects with known radioactivity concentrations and volumes) are frequently usedto validate segmentation and reconstruction techniques within nuclear medicine.Conventionally, uniform acrylic spheres are filled with radioactivity and insertedinto the standardized NEMA Image Quality (IQ) phantom (Fig. 1.1) to simulate tu-mours observed with PET. This allows for reproducible evaluation of scanner char-acteristics, but does not represent true patient anatomy or realistic tumour proper-ties.The aim of this project is to select the best reconstruction parameters and seg-mentation method for determining TMTV. For this, we will use highly sophisti-cated physical and simulated phantom experiments to model the bulky and hetero-geneous tumours commonly observed in PMBCL. PET images will be optimized5Figure 1.3: Overview of research study pipeline to evaluate improve quanti-tative metrics for PMBCL ensure accurate and reproducible determination of tumour burden for better pre-diction and management of PMBCL within a clinical setting. The ”roadmap” ofthis study is illustrated in Fig. 1.3 and will be discussed in following chapters.1.3 MotivationsIn essence, this thesis aims to design and implement novel techniques to evalu-ate research questions pertaining to Positron Emission Tomography (PET) withinoncology. Specifically, we aim to improve tumour quantification accuracy and ro-bustness within the context of metastatic prostate cancer and PMBCL. We selectedthese cancers due to the challenging image characteristics they present. Prostatecancer imaging with PSMA allows for smaller tumour metastasis to be detected,but quantitation accuracy can be significantly affected as lesions approach the res-olution limit of the PET scanner. Meanwhile, PMBCL quantitation represents avery different challenge within PET imaging. PMBCL characteristically presentswith very bulky and heterogeneous tumours, which necessitates validation of au-tomated segmentation methods. Within this thesis, our goal is to improve imagequantitation guidelines for prostate cancer and lymphoma, respectively, by apply-ing anthropomorphic phantom experiments.6Chapter 2Theory of Nuclear MedicineImagingThis section will identify the core physics principles pertaining to radiopharma-ceuticals, positron decay, reconstruction algorithms, and lesion segmentation tech-niques. All material within this section refers to Nuclear Medicine Physics: TheBasics, by Chandra and Rahmim [1], unless stated otherwise.2.1 Positron Emission TomographyIn positron emission tomography (PET), patients receive a specially-designed ra-diopharmaceutical (known as a tracer) to target different receptors or metabolicprocesses in cells. The efficacy of positron emission tomography (PET) relies onmaximizing tracer concentration in the target region (e.g. tumour), while min-imizing radioactivity in nearby tissue. Pharmacokinetic properties of the tracermay be modified to visualize different biological processes. For instance, 18F-fluorodopa is a PET tracer used to visualize dopamine nerve terminals and inves-tigate Parkinson’s disease, while 18F-fluorodeoxyglucose (18F-FDG) surveys forabnormal glucose-uptake and is commonly used in oncology [23]. All PET tracersare labelled with positron emitting radioactive isotopes, such as 18F, 68Ga, or 11C,in which a proton from the nucleus becomes a neutron, and a positron and neutrino7are ejected from the nucleus, as shown in equation 1:p+→ n+B++ ve, (2.1)When a positron interacts with an electron in tissue, an annihilation event oc-curs. Two 511 keV γ-rays are emitted in directions approximately 180◦ from eachother (Fig. 2.1.a).PET scanners contain detector rings optimized to detect 511 keV γ-rays. Twoevents measured within a short time-frame (e.g. in the order of nanoseconds) areassumed to have originated from the same annihilation event (Fig. 2.1.b). De-tector pairs from each perceived annihilation event are used to define the lines ofresponse (LORs). Projections, or views from the object at many different angles,are grouped into what is called a sinogram file. Sinograms contain the informa-tion needed to generate or reconstruct 3-dimensional images using analytical oriterative algorithms.2.2 Phantom Applications in PETThe NEMA Image Quality phantom (Fig. 2.2.a) is standardized and allows forreproducible evaluation of scanner performance, but does not realistically modelanatomy or heterogeneity within human PET imaging.Anthropomorphic phantoms, such as the Probe-IQ phantom (Fig. 2.2.b), rep-resents a significant advancement within phantom technology [24–26]. Probe-IQconsists of compartments designed to model a thorax (Radiology Support Devices,USA) and pelvis (Data Spectrum Corporation, USA). The thorax contains a fillableliver, lungs, ribs, and spinal cord, and is intended to model the size of a 92kg pa-tient [27]. Nylon mesh bags contain Styrofoam beads (Dow Chemical Corporation,USA) to displace water and establish realistic lung tissue density and radioactivity[26]. The Probe-IQ pelvis contains a 440mL compartment to simulate a bladder, aswell as 4mm clinical-grade tubing used to simulate ureters - relevant organs withinprostate cancer imaging, due to their high PSMA tracer uptake. Polyurethane filterfoam is used to position the organs within the phantom and create small pocketsof air bubbles within the foam to establish heterogeneous radioactivity concen-trations, which is more representative of a real human scan. However, physical8Figure 2.1: a.) Positron annihilation: Positron Emission Tomography(PET) tracers emit positrons, which travel through the surroundingmedium until they annihilate with an electron. Two 511 keV γ-raysare emitted from the annihilation location at approximately 180◦ direc-tions. Adapted from Chandra and Rahmim [1]. b.) Detector ring: PETscanners contain detector rings which detect 511 keV γ-rays. If two γ-rays are detected within a small time-frame then they are assumed tohave occurred from same annihilation event. Adapted from Uribe, 2019[2].9Figure 2.2: (a.) NEMA Image Quality (IQ) phantom. (b.) Probe-IQ phan-tom. (c.) Simulated 4D-extended cardiac torso (XCAT) phantom.anthropomorphic phantoms, such as Probe-IQ, are limited with respect to the factthat they can not easily model complex anatomical structures or be modified tostudy image characteristics at a population-scale. Thus, simulated phantoms havealso been developed for these purposes.The 4D-extended cardiac torso (XCAT) phantom allows for highly sophisti-cated multimodality imaging research (Fig. 2.2.c) [28, 29]. The XCAT phantomdefines dozens of organs, accounts for tissue-type (e.g. muscle, bone, soft tissue),and allows users to modify demographic features such as age, sex, and body massindex (BMI), providing the opportunity for phantom studies to be performed ata population scale. Since the phantom is simulated, the attenuation and radioac-tivity ground-truth is known at a sub-voxel level, allowing for more sophisticatedimaging metrics (e.g. shape, texture) to be evaluated. Due to the fact that thesephantoms are defined virtually, the PET image acquisition is not real and must alsoperformed through simulations. Therefore, it is very important to validate simula-tions through physical phantom experiments, to ensure that findings generalize toa real-life clinical setting.102.3 Reconstruction AlgorithmsOrdered Subset Expectation Maximization (OSEM) is currently the most prevalentalgorithm for PET image reconstruction [23, 30]. OSEM iteratively computes animage that has the maximum likelihood of generating the measured projections.The algorithm is derived from the maximum likelihood of the Poisson distribution.Reconstructions are computed with the following cost function:xˆ= argmaxx≥0nd∑i=1yilog[Px]i− [Px]i (2.2)where yi represents measured coincidence data, x is the image estimate, and Pis the system geometry matrix [30].However, a limitation of OSEM is that to achieve convergence it requires ahigh number of iterations, but each iteration increases the image noise. Because ofthis, the number of iterations are usually limited and convergence, especially forsmall objects, is usually not achieved. In practice, the algorithm is terminated after2-4 iterations [30].The Block Sequential Regularized Expectation Maximization (BSREM) algo-rithm adds a regularization term to the OSEM algorithm to remove the noise am-plification observed in OSEM [14, 30]. Noise is suppressed using the RelativeDifference Penalty (RDP), R(x), which is inserted into the cost function:xˆ= argmaxx≥0nd∑i=1yilog[Px]i− [Px]i−βR(x) (2.3)The RDP can be scaled using the β value (a smoothing factor), allowing usersto select their desired level of noise suppression. General Electric (GE) adver-tises their BSREM algorithm as Q.Clear [30] and it uses a default number of 25iterations; 10-13 times higher than regular OSEM. BSREM has the potential ofimproving image quality and quantification compared with OSEM [14, 30].PET images must also be corrected for degrading effects such as γ-ray scatter-ing, random event detection, and attenuation. For instance, computed tomographyattenuation correction (CTAC) scans are acquired to determine the attenuation fac-tors (µ) for the object. This allows for PET images to be corrected based on tissuedensity within the body.11PET image resolution is influenced by a variety of factors such as detectorsize, positron range (e.g. 0.2mm for 18F), annihilation angle, and inter-crystal blur-ring. Resolution blurring can be modelled using the point-spread function (PSF).PSF modelling is frequently applied to the OSEM algorithm, while the clinicalGE BSREM algorithm applies PSF by default. Although PSF can improve sig-nal recovery for small lesions, it may also result in SUV overestimation [24] andGibbs ringing artifacts. Modern scanners also incorporate time-of-flight (ToF),which detects picosend time-scale differences in coincidence events to localize theannihilation event with greater precision.2.4 Image SegmentationClinical reporting of tumour volume typically involves manual segmentation oflesions by a trained nuclear medicine physician (Fig. 2.3). This approach is con-sidered to be the gold-standard but is time-consuming and often suffers from highintra- and inter-observer variability [21]. Semi-automatic methods, including fixedthreshold (FT) and gradient-based segmentation, have been proposed and imple-mented within the literature [21].Fixed threshold algorithms allow for reproducible and efficient segmentationof tumours [21]. The European Association of Nuclear Medicine recommends a41% fixed-threshold, which selects all voxels with concentrations greater than 41%of the maximum uptake value within the tumour [31]. However, this method hasbeen observed to break-down for heterogeneous or low-uptake tumours [21].Gradient-based segmentation methods have shown recent advancements in le-sion segmentation tasks. A promising gradient-based approach is the PET Edge+tool by MIM (MIM Software, Inc.) which delineates tumour boundaries via inflec-tion points in the tumour concentration line profiles. Gradient-based methods arehighly vendor-specific and need to be carefully validated prior to introduction intoa clinical setting [21].2.5 Image QuantificationA variety of signal processing metrics have been applied within PET to comparedifferences in image quality. For instance, previous imaging studies have previ-12Figure 2.3: (Left-to-right) Manual, fixed threshold (FT), and gradient-basedsegmentation methods.ously correlated lesion detectability with the contrast-to-noise ratio (CNR) metric:CNR=CROI−Cbkgσbkg(2.4)where CROI and Cbkg are the measured concentrations in the region of inter-est and background, respectively. σbkg denotes the standard deviation of multiplerealizations of Cbkg.As referenced in Chapter 1, quantitative tumour metrics are of increasing im-portance within imaging, due to their relevance for outcome prediction tasks. Inthis thesis, we will compute lesion concentration accuracy using the recovery co-efficient (RC):RC =CmeasCtruth×100 (2.5)which compares the measured concentration, Cmeas, with its known concentra-tion,Ctruth. Concentration is typically normalized as the standardized uptake value(SUV):SUV =Cimg(t)ID/BW(2.6)Cimg(t) is the activity concentration decay-corrected to t = 0, which is normal-13ized by injected dose (ID) and patient body-weight (BW). The traditional imagingmetrics applied in clinical practice are derived from SUV and/or tumour volume:SUVmax Easiest to define and implement. Denotes the highest-concentration voxel in a ROI. It is highly vulnerable toPoisson noise and not considered to be a reproducibleimaging metric [32].SUVpeak Very robust and standardized method. Computed by find-ing the maximum 1mL spherical volume within a ROI.SUVmean Not standardized and segmentation-dependent. Defined asmean concentration within a ROI.TMTV Total volume of tumours (mL) as observed through a PETimage.TLG Total lesion glycolysis. Denotes the total observed radioac-tivity within the tumours.14Chapter 3Methods3.1 Quantitative Evaluation of Prostate Cancer3.1.1 PSMA Patient AnalysisMIM (MIM Software, USA) was used to analyze [18F]DCFPyL PET images of 10metastatic prostate cancer patients. The liver, lungs, bladder, ureters, mediastinum,abdomen, and pelvis were manually segmented. Organ-to-background radioactiv-ity ratios were computed. Lesions were manually segmented by a trained nuclearmedicine physician and SUVmean and MTV were determined.3.1.2 Shell-less 22Na LesionsThe long-lasting positron-emitter, 22Na, was used as an 18F analog to establishradioactivity within lesions models. 22Na has similar positron range and energyas 18F, allowing for realistic decay characteristics [33], but exhibits a significantlylonger half-life (2.6yrs versus 109.7min). Thus, 22Na has negligible radioactivedecay compared to [18F]FDG, allowing for variable lesion-to-background ratiosversus time (Fig. 3.1). Over 90 spherical lesions were cast using epoxy resininfused with [22Na]NaCl. Lesion concentrations (7-60 kBq/mL) and sphere diam-eters (3mm, 4mm, 5mm, 6mm, 7mm, 8mm, 10mm, 12mm, 14mm, 16mm) wereselected in order to achieve realistic SUVmean and MTV values determined from15Figure 3.1: Transaxial PET image slices of 22Na epoxy lesions in [18F]FDGbackground. Different lesion contrast is observed initially (t=0) com-pared to 4 hours later.Figure 3.2: (a.) Schematic of aluminum mold used for casting 3-16mmspheres. (b.) Radioactive epoxy spheres (3-16mm) infused with[22Na]NaCl.the PSMA patient analysis. A schematic of the aluminum mold used for the castingprocess, as well as the casted lesions can be observed in Fig. Probe-IQ Phantom Image AcquisitionThe highly realistic Probe-IQ phantom (Fig. 3.3), used in previous studies withsignificant impact on the field [26, 27, 34], was gifted to the Qurit Lab in 2019and with which extensive studies are ongoing [24]. The 22Na epoxy spheres wereinserted co-planar to each other in the Probe-IQ pelvis. The pelvis backgroundand bladder were injected with 65MBq and 140MBq [18F]FDG, respectively, toachieve target concentrations determined from the patient analysis. A dynamicPET/CT image of the Probe-IQ pelvis was acquired using the GE Discovery 690scanner (Fig. 3.4). A 2.5min frame duration was acquired when the total activityin the phantom approximated PSMA patient activity 1-hour post-injection. Bedduration of other frames were decay-corrected to maintain similar count statisticsthroughout the duration of the scan. An 18F scan with fully decayed background(3 days after initial acquisition) was used to measure the activity concentrations ofthe 22Na lesions.3.1.4 Data Processing and AnalysisThe Probe-IQ pelvis images were reconstructed using OSEM (24,32 subsets, 1-25iterations) and BSREM (32 subsets, 25 iterations, γ=2, β=0, 50, 100, 150, 200,250, 300, 400, 500, 650, 800). The BSREM algorithm was performed using a pro-prietary toolbox by GE Healthcare, referred to as Q.Clear. Point spread function(PSF) modelling is included with the BSREM algorithm and cannot be disabled inthe clinical scanner console. Therefore, for consistency, PSF modelling was ap-plied to both the OSEM and BSREM algorithms. No post-reconstruction filterswere applied.Lesion concentration was determined using SUVmax and SUVpeak, as well asSUVmean using a 40% of SUVmax fixed threshold (40% FT) and gradient-basedsegmentation algorithm, PET Edge+ (MIM Software, USA). In this work, we alsodefine a new metric denoted as SUVapex. We defined SUVapex as the mean concen-tration of a 6-voxel ROI (0.26mL) centered at SUVmax.The ground truth of total lesion activity was determined by placing a largespherical ROI over each lesion to compensate for the spill-out effect. The volumeof each sphere was known through the casting process, which allowed for lesion17Figure 3.3: (Top) Medium and large phantom shell. (Bottom) Pelvic phantomcompartment with bladder insert, as well as ribs and spine that can beinserted in the large phantom.activity concentration to be calculated. Recovery coefficients (RC’s) were plottedfor each sphere, after applying each combination of reconstruction parameters andsegmentation methods.3.1.5 Canadian PET Prostate Phantom for OncologyGiven the increasing number of PSMA clinical trials in North America, there issignificant motivation to improve generalizability between different scanners andcancer centres. Although the Probe-IQ phantom is state-of-the-art within its field,the phantom filling protocol is extremely time-consuming and users typically re-ceive greater radiation dose compared with conventional phantom scans. Thus, theProbe-IQ phantom is primarily restricted to in-house imaging studies.In response, we introduce the Canadian PET Prostate Phantom for Oncology(C3PO) - a PET/CT/MRI-compatible phantom for scanner harmonization. The18Figure 3.4: (a.) PET image of metastatic prostate cancer patient imagedwith [18F]DCFPyL, targeting Prostate-Specific Membrane Antigen. (b.)PET image of Probe-IQ phantom.phantom was designed through a collaboration with the BC Cancer machine shop.C3PO contains acrylic cylinders with hollowed out spheres, ranging from 3-37mmin diameter. [18F]FDG can be injected into each acrylic sphere to simulate lesionswith ideal imaging conditions (e.g. without background). The C3PO cavity fea-tures a centrally-located 30mL vial to simulate a bladder, as well as sodium-22epoxy spheres (Fig. 3.2) to simulate prostate cancer metastasis. [18F]FDG can beinjected into the cavity to simulate background tracer uptake. Polyurethane filterfoam is used to position the epoxy spheres and bladder. The foam also createssmall pockets of air, thus displacing radioactivity and establishing heterogeneity inthe PET images. C3PO blueprints are attached in the Appendix (A.1, A.2, A.3).C3PO Study ProtocolSodium-22 epoxy lesions (3, 5, 6, 7, 8, 10, 14, 16mm diameters) were insertedco-planar into the prostate phantom. [18F]FDG was injected into the bladder andbackground regions to achieve target concentrations (120 kBq/mL and 2 kBq/mL,19Figure 3.5: Canadian PET Prostate Phantom for Oncology. (a.) Externalview of phantom. (b.) View of opened phantom to show epoxy lesions.(c.) Acrylic cylinders that can be inserted into phantom.respectively) determined from the PSMA patient analysis. Bladder and backgroundtarget volumes were determined by weighing the phantom compartments pre andpost-filling. 6 acrylic cylinders (3mm, 7mm, 13mm, 17mm, 25mm, 37mm) wereinjected with [18F]FDG to achieve 60 kBq/mL activity concentrations. An 18Fscan with fully decayed background was used to measure radioactivity of the 22Nalesions. A study protocol is shown in the Appendix (A1).Images were acquired using the GE Discovery 690 and GE MI PET/CT scannerwith time-of-flight enabled. Images were reconstructed with OSEM (4 iterations, 8subsets) with PSF. Post-reconstruction transaxial filters were applied to the nativeimage (2mm, 4mm, 6mm, 8mm). Lesions generated from 22Na epoxy spheres andacrylic cylinders were segmented using 40% FT and gradient-based algorithms.20TLG ground truth was determined by placing a large spherical ROI over lesionsin the image with fully-decayed background. MTV ground truth was determinedby the known volume of each sphere in the mold. Recovery coefficients werecalculated for each lesion.3.2 Non-Hodgkin’s Lymphoma Quantification3.2.1 Physical Phantom ExperimentsPMBCL Patient AnalysisMIM (MIM Software, USA) was used to analyze ten [18F]FDG PET/CT scansof PMBCL patients. Regions of interest (ROIs) including the liver, lungs, medi-astinum, abdomen, and pelvis were be manually segmented and organ-to-backgroundactivity ratios were computed. An expert nuclear medicine physician detected andsegmented the bulkiest tumour for each patient using a gradient-based segmen-tation algorithm and SUVmean and MTV were determined. Lesion segmentationswere saved in a stereolithography (stl) file and edge-smoothing effects were applied(Fig. 3.6.a).Novel Tumour Casting MethodFive lesions ranging from 2.7mL to 76mL were 3D-printed. Aluminum rods wereembedded in the 3D-printed models to simplify the casting process. Negativesof the 3D-printed tumours were casted using silicone-based molding materials(Smooth-On, USA). Tumour models were casted using a mixture of liquid plas-tic (Smooth-On, USA), fluorescent red pigment, and [18F]FDG (Fig. 3.6.b). The[18F]FDG concentration (21.5 kBq/mL) corresponded to the 50th percentile of 22lesions from a cohort of 8 PMBCL patients. The red pigment was used to ensureuniform mixing of the [18F]FDG. The liquid plastic mixture was injected into thenegative-cast silicone molds. MTV ground truth was determined using lesion massand density of liquid plastic (ρ=1.15g/mL). Tumour models were removed fromthe molds and inserted into the mediastinum of the Probe-IQ phantom.21Figure 3.6: Probe-IQ lymphoma experiment pipeline. (a.) Segmentation ofbulky PMBCL tumour saved in a stereolithography file format. (b.) Ra-dioactive PMBCL tumour model casted with liquid plastic. (c.) PET/CTcoronal slice of PMBCL tumour model inserted into Probe-IQ thoraxshell.Image AcquisitionTumour models were inserted into Probe-IQ phantom without any background ac-tivity. 2x10min bed positions were acquired using the GE MI PET/CT scanner(Fig. 3.6.c). This acquisition was used to measure the total activity of the [18F]FDGlesions and obtain the ground truth for radiomics features (e.g. shape, texture).Next, the Probe-IQ liver and thorax were injected with [18F]FDG to achieve tar-get concentrations determined from the PMBCL patient analysis. 2x30min bedpositions were acquired in list-mode.Data Processing and AnalysisData was unlisted into 2x1min bed durations. The images were reconstructed us-ing OSEM (4 iterations, 8 subsets) with time-of-flight (ToF) and point-spread func-tion modelling (PSF), with a 256x256 matrix size. Images were segmented usingMIM (MIM Software, USA) with 20%, 25%, 30%, 40%, 50% FT and MIM’s PETEdge+ gradient-based algorithm. Total metabolic tumour volume (TMTV) and to-tal lesion glycolysis (TLG) were calculated for each of the tumours.22Table 3.1: Measured [18F]FDG activity concentrations for PMBCL patientsin different regions of interest.Region of Interest Concentration (kBq/mL)Background 1.5Bladder 38.5Esophagus (outer, contents) 4.1, 3.6Heart (bloodpool, myocardium) 11.6, 55.0Intestines 6.7Kidney (medulla, cortex, pelvis) 19.8, 8.4, 25.3Liver 13.1Lung 3.5Spleen Simulated Phantom ExperimentsUpgrades to XCAT PhantomThrough collaboration with Paul Segars (Duke University), the 4D-extended car-diact torso (XCAT) phantom was upgraded (Fig. 3.7.a). A template lymphaticsystem model based on anatomical data from the Visible Human Project of theNational Library of Medicine was used to define 276 lymph nodes and corre-sponding vessels using non-uniform rational basis spline (NURBS) surfaces (Fig.3.7.b). The multichannel large deformation diffeomorphic metric mapping (MC-LDDMM) method was used to propagate from the template phantom to differentXCAT anatomies. This allows for the lymphatic system to be investigated on pa-tients with different genders, weight, sizes, age, and other anatomical differences.Lymph node properties were modified using the Rhinoceros 3D viewing software.The XCAT general parameter script was used to input organ concentrations andgenerate binary files with uptake and attenuation information (Fig. 3.7.c). Uptakeparameters were set based on an analysis of 5 [18F]FDG PET/CT images of patientswith PMBCL (Table 3.1). The phantom was used as the input to a MATLAB-basedPET simulation and reconstruction tool [35] generating simulated PET/CT imagesfor a GE Discovery RX scanner, reconstructed with OSEM (2 iterations, 24 sub-sets).23Figure 3.7: Lymphatic system development pipeline. (a.) Anterior view ofmale (left) and female (right) 4D XCAT phantom anatomies. Adaptedfrom figure by Segars et al., 2008. (b.) Lymphatic system incorporatedfor XCAT phantom, which can be combined with XCAT organs. (c.)Coronal slice of combined XCAT and lymphatic phantom showing ra-dioactivity and attenuation distribution for a ”healthy” patient.Tumour SimulationTen PMBCL patients with heterogeneous lymph node conglomerates were simu-lated using the XCAT phantom with an integrated lymphatic system. Organ activ-ity concentrations were specified based on analysis of [18F]FDG PET/CT imagesfrom 5 PMBCL patients. Tumours with 4-100mL volumes were generated withactivity concentrations of 21.5 kBq/mL. The concentrations corresponded to the50th percentile of lesions segmented in the PMBCL patient analysis. To createbulky lymph node conglomerates in the XCAT phantom, lymph node morphologyand function was altered: lymph nodes were expanded, stretched asymmetrically,converged within the mediastium, and heterogeneous lymph node activity was es-tablished (Fig. 3.8).Data AnalysisRegions-of-interest were drawn using MIM (MIM Software Inc.) using 20%,25%, 30%, 40%, 50% FT and MIM’s PET Edge+ gradient-based algorithm. Totalmetabolic tumour volume (TMTV) and total lesion glycolysis (TLG) were calcu-lated for each of the tumours. TMTV and TLG percent bias were plotted versusthe ground truth.24Figure 3.8: PMBCL tumour simulation. (Left) Lymph nodes expanded andconverged within Rhinoceros 3D viewing software, with additional or-gans shown. (Right) Coronal slice of radioactivity and attenuation filegenerated for patient with bulky mediastinal tumour.25Chapter 4Results4.1 PSMA Experiments4.1.1 Lesion SegmentationsFocal, high contrast lesions (3mm-16mm) were segmented and visualized in Fig.4.1. 3 metrics with fixed-volumes were applied: SUVmax, SUVpeak, and SUVapex.SUVmax selects the highest concentration voxel in the lesion. SUVpeak utilizesan appropriate volume for the 14mm lesion, but appears to be too large for ≤8mmlesions. SUVapex defines a volume that is intermediate to the previous two methods(0.26mL or 6 voxels), which corresponds to the size of an 8mm sphere.The fixed threshold (FT) segmentation selects a % of SUVmax, and so a jaggedcontour is observed due to its voxel-by-voxel approach. Qualitatively, the segmen-tation appears to significantly overestimate the the 6mm and 8mm lesion. Mean-while, the gradient-based method smoothly follows the boundary of the 6mm,8mm, and 14mm lesions.4.1.2 Recovery CurvesLesion recovery coefficients (RC’s) were plotted for different reconstruction algo-rithms and segmentation methods (Fig. 4.2). PSF modelling resulted in overesti-mated lesion concentration using the SUVmax metric (Max RC). This overestima-26Figure 4.1: Segmentation methods applied to PET images of 14mm, 8mm,and 6mm 22Na epoxy spheres.tion peaked for the 10mm diameter lesion, and was amplified for higher iterationsfor OSEM (294.8% for 32 subsets, 5 iterations), and lower β values for BSREM(293.9% for β=50). Using BSREM, RC’s were significantly underestimated for le-sions less than 10mm, and β=100 was required to minimize recovery loss throughsignal smoothing (25.4% and 77.9% for 5mm and 7mm respectively). For OSEM,at least 3 iterations was required to increase concentration recovery (114.9% and91.4% for 5mm and 7mm respectively).The SUVpeak recovery curve followed a monotonic, increasing relationshipwith respect to lesion diameter. For OSEM (32 subsets, 2 iterations), the 8, 10,12, 16mm lesions had 23.2, 54.6, 71.5, and 118.7% recovery. Meanwhile, forBSREM (β=300), the same spheres had recovery coefficients of 18.7, 43.2, 58.7,and 99.3%.SUVapex recovery curves sharply increased from 6-10mm, and plateaued in the10-12mm range. The 10mm and 12mm were most accurate using BSREM β=100(89.9% and 97.8% respectively), and OSEM with 2 iterations (99.3% and 105.3%,using 32 subsets).27SUVmean with fixed threshold was most accurate using BSREM β=200 (88.4%,91.9%, and 108.7% for 10mm, 12mm, and 16mm respectively). Meanwhile, lowerβ values minimized signal loss for smaller lesions (12.3%, 24.5%, and 64.1% for4mm, 6mm, and 8mm respectively, using β=100). For OSEM, the most accuraterecovery for larger lesions was achieved with 1 iteration (75.7, 101.5, 118.0% for10mm, 12mm, and 16mm). Higher numbers of iterations (4 or 5) are needed tominimize signal loss for smaller lesions (4-8mm).SUVmean with the gradient method was most accurate for larger lesions usingBSREM β=300 (69.5%, 89.2%, and 121.9% for 10mm, 12mm, and 16mm respec-tively). β=50 was necessary to increase recovery for smaller lesions (17.1, 45.0,74.7% for 4mm, 6mm, and 8mm respectively). OSEM with 32 subsets was mostaccurate for 12-16mm lesions after 1 iteration (84.2, 100.9, 147.6% for 10mm,12mm, and 16mm respectively). Smaller lesions were underestimated with OSEM,so higher iterations are preferred in order to maximize the signal.SUVpeak had lower mean absolute error (MAE) than SUVmax. SUVapex, a hy-brid between SUVmax and SUVpeak, significantly reduced MAE for both recon-struction algorithms (OSEM and BSREM were 55.1±51.4 and 60.9±42.6, respec-tively). Both the fixed threshold and gradient-based segmentation methods hadsimilar MAE compared to SUVapex.4.1.3 Robustness of Recovery CurvesThe recovery curves were plotted versus lesion-to-background ratio for the 10mmlesion (Fig. 4.3). The pink shaded regions in the figure are used to highlight thedomain with background concentrations that are observed in PSMA PET images.Qualitatively, SUVmax and SUVmean FT appear to be the least stable versus lesion-to-background ratio. This is validated quantitatively, as SUVmax and SUVmean(FT) had the largest range and standard deviation, respectively. The gradient-basedSUVmean had significantly lower range and standard deviation, and appeared to bemore stable versus lesion-to-background ratio. SUVpeak exhibited the lowest stan-dard deviation and range, while SUVapex had the second-best values. However, itshould also be noted that SUVpeak underestimated recovery for this lesion diame-ter (10mm). SUVapex appears to have the best combined accuracy and precision,28Figure 4.2: (Top to bottom) Recovery concentration coefficients measured inProbe-IQ pelvis using Max, Peak, Apex, and Mean (40% FT and gradi-ent). (Left to right) Reconstruction algorithms using OSEM+PSF) andBSREM (32 subsets). Mean Absolute Error (MAE) +/- Standard Devi-ation indicated on each plot. 29compared with the other methods.4.2 Lymphoma Quantification4.2.1 Physical Phantom ExperimentsThe PMBCL tumour models were inserted into the Probe-IQ phantom. PET im-ages were acquired in list-mode. 1min bed-durations were unlisted and recon-structed. Lesions segmented with fixed threshold (FT) and gradient-based methodsare shown in Fig. 4.4. Qualitatively, the 20% threshold is not desirable for segmen-tation as it included voxels within the background. The 25% and 30% thresholdsdelineated the entire tumour without spilling into background. Thresholds ≥40%neglected significant regions of the tumour. The gradient method smoothly fol-lowed the boundary of the lesion, and segmented smaller volumes than the 25-30%thresholds.TMTV percent bias versus ground truth is shown in Fig. 4.5. TMTV percentbias, obtained with a 25% FT, for each lesion was -7.6% (3mL), -14.5% (21mL),and 2.3% (71mL). Percent bias in TMTV using PET Edge+ was -3.5% (3mL),-32.5% (21mL), and -24% (71mL).TLG percent bias using the 25% FT had a percent bias of -18.3% for the 3mLlesion and 6.6% for the 71mL lesion. TLG percent bias with PET Edge+ was-17.3% for the 3mL lesion and -11.6mL for the 71mL lesion.4.2.2 Simulated Phantom ExperimentsTumour SimulationThe lymphatic system was added to the XCAT phantom with the capability toselect male/female anatomy, as well as to modify demographic features such aspatient height, weight, or body mass index (BMI). Lymph nodes can be scaled,asymmetrically stretched, and translated within the intuitive Rhinoceros interface,to allow for realistic simulation of different lymph node pathologies. Using thisadaptable lymphatic system, lymph node conglomerates with any desired anatomi-cal formation or tracer uptake can be simulated. As a proof-of-concept, the specific30Figure 4.3: Recovery concentration coefficient versus lesion-to-backgroundratio for 10mm lesion. (Top to bottom) Max, Peak, Apex, and Mean(40% FT and gradient). (Left to right) Reconstruction algorithms usingOSEM+PSF and BSREM (32 subsets). Range and standard deviation ofrecovery coefficients annotated on plots. Realistic PSMA backgroundactivity concentration denoted by pink shaded region.31Figure 4.4: PET image of PMBCL tumour model inserted into Probe-IQphantom. Tumour segmented using 20%, 25%, 30%, 40%, 50% fixedthreshold (FT) and gradient-based PET Edge+ method.anatomy of a tumour in a PMBCL patient was replicated (Fig. 4.7). The gener-ated [18F]FDG PET images of PMBCL patients were assessed to be realistic by anexperienced nuclear medicine physician. This novel software upgrade and appli-cation using the XCAT phantom represents a significant advancement in tumourgeneration technology.Segmentation ResultsBulky, heterogeneous PMBCL tumours, ranging from 4mL to 100mL, were gen-erated in the mediastinum of the XCAT phantom using the previously discussedtumour simulation method. Fixed threshold (FT) and gradient-based segmenta-tions are shown in Fig. 4.7. The gradient-based method smoothly follows the32Figure 4.5: Percent bias vs. true value plotted for: (Top) Total metabolictumour volume (TMTV) and (Bottom) Total lesion glycolysis (TLG)plotted versus true value, using PMBCL tumour models casted using[18F]FDG-infused liquid plastic, with 1 minute bed duration.33Figure 4.6: Coronal views of PET images using MIM. Bottom images in-clude gradient-based segmentation of tumour. (Left) [18F]FDG PETimage of PMBCL patient with bulky mediastinal tumour. (Right) Sim-ulated PET image using XCAT phantom. Tumour consisted of 5 ex-panded lymph nodes with heterogeneous activity concentrations.34curvature of the lesion and sharp corners are not observed. For comparison, the FTmethod takes a % of SUVmax approach, and so jagged contours are observed due tothe voxel-by-voxel nature of the segmentation method. Certain voxels within thetumour appear to be neglected for thresholds≥40%.TMTV percent bias versus ground truth is shown (Fig. 4.8). TMTV percentbias, obtained with a 25% fixed threshold, for the different tumour volumes was4.4% (13mL), -0.7% (39mL), -11.4% (71mL), and -14.7% (100mL). Percent biasin TMTV with PET Edge+ was 22.1% (13 mL), 11.9% (39 mL), 26.3% (71 mL),and 20.5% (100 mL).TLG percent bias versus ground truth was also plotted (Fig. 4.8). The 25%threshold led to percent bias of -15.6% for the 100mL lesion and -8.5% at 13mL.PET Edge+ TLG percent bias was -5.0% at 13mL, and only deviated -1.6% for the100mL lesion.35Figure 4.7: PET image of PMBCL tumour simulated using XCAT phantom.Tumour segmented using 20%, 25%, 30%, 40%, 50% fixed threshold(FT) and gradient-based PET Edge+ method.36Figure 4.8: Percent bias vs. true value plotted for: (Top) Total metabolictumour volume (TMTV) and (Bottom) Total lesion glycolysis (TLG)plotted versus true value. Lesions consisted of 2-4 asymmetric lymphnodes converged within the mediastinum, and ranged from 4-100mL.37Chapter 5Discussion5.1 Applications to Prostate Cancer5.1.1 PSMA Lesion QuantificationRecovery curves for SUVmax, SUVpeak, SUVapex, SUVmean (40% FT and gradient)are shown in Fig. 4.2. SUVmax RC’s were significantly overestimated for spherediameters close to 10mm. The peak at 10mm was most certainly caused by PSFmodelling [36]. Selecting higher β values (BSREM) or reconstructing for less it-erations (OSEM) minimized overestimation of RC’s. The SUVmax metric is notconsistent versus lesion-to-background ratio, as shown in Fig. 4.3. This is antic-ipated, since SUVmax depends on a single-voxel and can be greatly influenced bynoise [32, 36]. Therefore, we can conclude that SUVmax does not appear to be anappropriate metric for quantification of PSMA PET images.The RC overestimation at 10mm, present in the SUVmax curve, was not ob-served using SUVpeak. In fact, SUVpeak underestimated the RC of all lesions≤12mm diameter. Plotting RC versus lesion-to-background for the 10mm lesion,it had standard deviations of only 6.6% and 12% for the OSEM and BSREM al-gorithms, respectively. Therefore, SUVpeak appears to be the most robust metricversus different lesion-to-background ratios (Fig. 4.3). Although SUVpeak is notaccurate for lesions ≤12mm, it should still be considered for use within clinicalpractice due to its robustness.38Our newly-defined metric, SUVapex, resulted in intermediate RC’s comparedwith the SUVmax and SUVpeak methods. SUVapex appears to be accurate for 10-16mm lesions, given that a reasonable selection of β (≈ 100-400)is applied forBSREM. RC overestimation for the 10mm lesion, present with SUVmax, was notobserved using SUVapex. Additionally, 8-10mm RC’s were not as severely under-estimated, compared with SUVpeak. In terms of robustness, SUVapex was quiteconsistent versus the lesion-to-background ratio. It had standard deviations of28.5% and 33.8% for the OSEM and BSREM algorithms, respectively. Thesevalues are higher than SUVpeak, but much lower than those observed for SUVmax.Therefore, SUVapex appears to be a possible ”happy medium” between SUVmaxand SUVpeak: it increases RC’s for smaller lesions, but minimizes RC overesti-mation due to PSF and appears to be a reasonably robust metric. Further researchneeds to be performed to evaluate different variations of SUVpeak for PSMA quan-tification. These metrics may involve different contour sizes, shapes (e.g. sphericalversus circular), and localization (e.g. centered on SUVmax versus finding the high-est uptake) [37].SUVmean RC’s were evaluated using the 40% fixed threshold (FT) and gradient-based segmentation methods. SUVmean using 40% FT appears to be most accuratefor β=200-400 with BSREM, or 1 iteration with OSEM. However, as observed vi-sually, OSEM images with only 1 iteration are not converged, so performing recon-structions with this parameter is not recommended. Similar to SUVmax, RC’s weresignificantly overestimated for 10mm lesions for certain reconstruction parame-ters. This is anticipated, since this method thresholds based on SUVmax, whichwas overestimated due to PSF modelling. As shown in Fig. 4.3, RC’s using 40%FT were highly variable versus lesion-to-background ratio. In fact, standard de-viation (141.5% and 114.3%, for OSEM and BSREM, respectively), were nearlyas high as the values observed using SUVmax. As a result, the 40% FT does notappear to be very robust, and is not recommended for quantification of PSMA PETlesions.SUVmean RC’s determined using the gradient segmentation are shown in Fig.4.2. The gradient method was more robust at minimizing the recovery peak at10mm, compared with the 40% FT. Given a proper selection of reconstructionparameters (e.g. β ≈ 200-400), gradient and FT segmentation methods appear39to have comparable accuracy for recovery coefficients. However, as shown byplotting RC versus lesion-to-background ratio (Fig. 4.3), the gradient method hadsignificantly lower standard deviation (59.9% and 58.8% for OSEM and BSREM,respectively). Due to its consistency, gradient-based segmentation appears to bemore appropriate than the 40% FT for SUVmean quantification of PSMA tumours.5.1.2 PSMA PET Harmonization StudyIt is our goal to expand the Canadian PET Prostate Phantom for Oncology study tocentres across North America and Europe. Currently, we have recruited approxi-mately 12 nuclear medicine departments for the initial phase of study. The C3POscan will be replicated at each site and image-processing methods (e.g. post-filtersmoothing, modifying number of reconstruction iterations) will be applied to har-monize lesion quantification. This study is extremely promising, as harmonizedPSMA PET images will lead to improved data sets for prostate cancer images,and ideally lead to the development of robust outcome prediction models (e.g. pa-tient survival, disease progression) for prostate cancer. In the big picture, we hopethat research advancements resulting from C3PO will allow physicians to generatemore accurate prognosis of prostate cancer.5.2 Applications to Lymphoma5.2.1 Lymphoma SimulationWe collaborated with Dr. Paul Segars (Duke University) to upgrade the XCATphantom to contain a fully-defined lymphatic system. Using the radioactivity andattenuation ground truth generated from the XCAT phantom, we were able to gen-erate realistic PET/CT images. Intended for open-source usage, this upgradedXCAT phantom has significant potential to improve image quality and quantifica-tion studies related to lymphoma. In fact, we used the lymphatic system to definerealistic PMBCL tumours. These simulated tumours were determined to be highlyrealistic by an experienced nuclear medicine physician. We simulated 10 patientswith lymphoma tumours and evaluated FT and gradient-based segmentation meth-ods.40As shown by visual comparison of segmentation methods (Fig. 4.7), selectinga 20-30% FT appears to result in accurate delineation of the tumour. This wasconfirmed quantitatively in Fig. 4.8. TMTV quantification using 20% FT wasmost accurate for lesions ≤50mL, while the 25% FT was best for lesions ≥50mL.All FT segmentations underestimated the total lesion glycolysis. TLG percent error20-30% FT’s was within 25%. The 40% and 50% thresholds typically had percentbias greater than 25%. Therefore, FT’s between 20% and 30% appear to be mostsuitable for TMTV and TLG quantitation.On average, the gradient-based method delineated larger regions than the 20%FT. In fact, the gradient-based method consistently overestimated TMTV (Fig.4.8). However, the gradient-based segmentation method was more accurate thanthe FT methods for determining TLG. This is likely because the overestimatedboundaries compensates for the spill-out effect. Overall, our simulations indicatethat the ≈25% FT and gradient-method are most accurate for TMTV and TLGquantitation, respectively.5.2.2 Physical Phantom ExperimentWithin this thesis, we developed Negative-Cast Modelling for Oncology (NCMO),a template-based approach for accurate and reproducible casting of tumour mod-els. We applied NCMO to validate PMBCL quantitation results from the XCATsimulations. As shown in Fig. 4.4, the 20% FT overestimates tumour boundaryand selects voxels located in the background. Conversely, the 40-50% FT failed tosegment certain regions within the tumour. Therefore, 25-30% FT appears to bemost accurate from a qualitative perspective. TMTV quantitation was most accu-rate using the 25% FT, and significantly overestimated using 20% FT. These resultsare reasonably consistent with the XCAT simulation.The gradient-based method failed to segment certain regions of the tumour andunderestimated TMTV. As described in 2, the gradient method segments regionsby finding inflection points in the tumour concentration line profiles. The PMBCLtumour models had heterogeneous radioactivity distributions, which likely resultedin the segmentation method detecting local minima within the tumour. Furtherinvestigation is required to assess whether PET Edge+ is a robust tool for PMBCL41segmentation tasks.As validated through physical and simulated phantom experiments, the 25%FT appears to result in the most accurate tumour quantitation for PMBCL tumours.This encourages further application of this segmentation method. For example, infuture studies we may apply the 25% FT to delineate the ground truth for super-vised machine learning algorithms. Overall, our goal is to introduce routine PM-BCL segmentation into clinical practice, which will hopefully lead to improvedoutcome prediction and disease management of lymphoma.5.3 Advancements in Phantom TechnologyThis thesis culminated in multiple phantom imaging developments at BC Cancer.Fig. 5.1(a-d) illustrates the variety of PET images that were generated using phan-tom technology available within the Qurit Lab. As shown in Fig. 5.1.a, the NEMAImage Quality phantom simulates spherical tumours within an abdomen-shapedcompartment. This phantom is very useful for performing studies relating to im-age quality. For instance, it is frequently used to select reconstruction parametersin order to improve lesion contrast. However, the NEMA IQ phantom does notmodel organs that are highly relevant within PET (e.g. liver or bladder). It alsouses fillable spheres to model lesions, which results in lesion concentration andvolume biases due to the cold-shell effect [11, 12]. Therefore, the NEMA phantomwas not desirable for evaluating quantitative imaging metrics.The Probe-IQ phantom (Fig. 5.1.c) was the first advancement within this thesistowards anthropomorphic phantom imaging. The Probe-IQ phantom approximatesthe anatomy of major organs, such as the liver and lungs, allowing for realistic ra-dioactivity distributions to be established within the phantom. It contains porousfilter foam, which serves two important purposes. Firstly, the foam creates smallair pockets to displace radioactivity and increase image heterogeneity, which ismore realistic to a patient image. Secondly, it secures the organ inserts and allowsus to embed tumour models within the phantom. To evaluate lymphoma, we devel-oped a novel tumour casting method to generate realistic tumour models. Withinthe context of prostate cancer, we developed shell-less radioactive epoxy spheres,which simulates lesions imaged with PSMA PET. Although the Probe-IQ phantom42experiments were effective at evaluating PET quantitation, it should be noted thatthese experiments can also be extremely arduous and time-consuming. They typi-cally require multiple months of preparation and often result in increased radiationdose to its users.In response to the challenges described above, we developed the Canadian PETProstate Phantom for Oncology (Fig. 5.1.b). C3PO represents a ”middle-ground”between the Probe-IQ and the NEMA phantom. By utilizing 22Na epoxy lesions,it removes the cold-shell effect that is present within the NEMA phantom. It alsosimulates basic anatomy such as a bladder, which is highly relevant to PSMA PETimaging. However, compared to the Probe-IQ phantom, the filling procedure isgreatly simplified and is very easy to replicate by different users. Additionally,the total activity of the epoxy spheres is less than 1MBq (due to their small sizes),which allows it to be shipped under an ”exempt” status by the Canadian NuclearSafety Comission (CNSC). The protocol requires less than 60min for preparationand only 9min scan duration, making it feasible for large-scale imaging studies. Byproviding a simple solution for PSMA PET harmonization, C3PO has the poten-tial to improve comparison between clinical trials which may lead to more robustpredictive metrics and prognosis for prostate cancer.One disadvantage to physical phantoms is that they are not easily scaled torepresent patients with different demographic features. Conversely, the simulatedXCAT phantom (Fig. 5.1.d), can be easily modified to account for features such aspatient sex, height, and weight. Composed of different tisuse types and dozens oforgans, XCAT is significantly more sophisticated than physical phantoms. Throughdevelopments described within this thesis, the XCAT phantom also incorporates276 lymph nodes and corresponding vessels.Finally, for comparison, Fig. 5.1.e shows a PMBCL patient imaged with[18F]FDG PET. By visual comparison, XCAT appears to be the most realistic phan-tom available to us. These simulated phantoms with enhanced realism create thepotential for more sophisticated phantom studies. However, it must be met withcaution, as these acquired images are not ”real” and so it is very important that wevalidate our experiments with physical phantoms.43Figure 5.1: (a.) Conventional NEMA Image Quality phantom. (b.) Cana-dian PET Prostate Phantom for Oncology. (c.) Image quality probe(Probe-IQ) phantom. (d.) Simulated 4D-extended cardiac torso phan-tom (XCAT). (e.) Non-Hodgkin’s lymphoma patient imaged with PET.5.4 Future Directions5.4.1 Dosimetry for Radiopharmaceutical Therapy[177Lu]-PSMA-617 to treat metastatic prostate cancer is currently undergoing phaseIII clinical trials at BC Cancer. However, under the current protocol, every pa-tient receives the same injected activity, regardless of the patient’s metabolism or44body mass. In the future, we hope to introduce personalized dosimetry for [177Lu]-PSMA-617 therapy using SPECT/CT imaging. Therefore, it is of high importancethat we evaluate the accuracy of SPECT imaging quantiation prior to its introduc-tion into the clinic.Our intention is to modify the Probe-IQ phantom experiment to contain 177Lu.Lesions will be cast using the 75Se isotope, which acts as a long-lasting analog to177Lu. Clinical parameters (e.g. scan time, reconstruction and segmentation meth-ods) will be modified to ensure accurate and reproducible delineation of traditionaltumour metrics (e.g. tumour volume). We will also use compute radiomics fea-tures (e.g. texture), using high-resolution images from a small-animal scanner todetermine the ground-truth.Next, the 4D extended cardiac-torso (XCAT) phantom will be implemented forprecise, voxelized dosimetry of PSMA SPECT/CT. The XCAT phantom – renownfor its realistic anatomical structures – will be used to define the ground truth ra-dioactivity and attenuation for a simulated patient. Focal, high-contrast PSMAlesions will be defined using the XCAT general parameter script. Next, we willuse SIMIND Monte Carlo to generate simulated SPECT/CT images of the XCATphantom. Voxelized dosimetry will be applied to determine the best protocol forcomputing dose to healthy and malignant tissue. Radiomics features will alsobe evaluated, in order to investigate the feasibility of radiomics-based dosimetrywithin a clinical setting.Overall, these phantom studies will allow for quantitative evaluation of organand voxel-level dosimetry methods. Ideally, these experiments will help us selectthe best clinical protocols for use within the clinic, and lead to better clinical out-comes for radiopharmaceutical therapy.5.4.2 Outcome Prediction for Non-Hodgkin’s LymphomaThis thesis culminated in the development of a novel tumour casting method. Thisallows us to cast tumour models of any desired size or shape. However, this proce-dure is somewhat time-consuming and can be challenging from a radiation-safetyperspective. In the future, we hope to retrofit a 3D-printer to directly print withradioactive materials, as recently described by Gear et al., 2020 [38]. This would45allow for enhanced accuracy and precision while casting complex tumour struc-tures.Furthermore, we intend to implement this tumour casting method to identifyrobust radiomics features. For example, we could evaluate which metrics are con-sistent for different image contrast and noise realizations. Ideally, this will lead tobetter feature selection for machine learning models, which may lead to improvedoutcome prediction for lymphoma patients within a clinical setting.46Chapter 6Conclusions6.1 PSMA PET QuantificationWithin this thesis, we found that image reconstruction using BSREM with β=200-400 (32 subsets) or OSEM with 1-2 iterations (32 subsets) resulted in the mostaccurate tumour quantification. For PSMA lesion segmentation, the gradient-basedmethod is recommended over the 40% fixed threshold because it is more robustand less-dependent on noise. SUVapex, a hybrid method of SUVmax and SUVpeakwas proposed as a segmentation-free solution for quantification of PSMA lesions.SUVapex is more robust than SUVmax, but features improved accuracy compared tothe conventional SUVpeak metric.The Canadian PET Prostate Phantom for Oncology (C3PO), a PET/CT/MRI-compatible phantom, was developed for harmonization of PSMA PET imaging.C3PO allows for efficient and reproducible modelling of the focal, high contrastlesions that are frequently observed for metastatic prostate cancer. Intended to bedeployed within a large-scale study, C3PO has significant potential for enablingimproved PSMA PET harmonization.6.2 PMBCL QuantificationThe highly sophisticated XCAT phantom was upgraded to contain a lymphatic sys-tem. We successfully simulated realistic PET/CT images of PMBCL patients by47modelling bulky lymph node conglomerates within the mediastinum. Made pub-licly available, the XCAT phantom with the new lymphatic system has the poten-tial to improve image quality and quantification experiments, towards improvedassessment of lymphoma such as with predictive modelling.To validate these simulations, we developed a negative-cast molding techniquefor use in oncology. Using this method, physicists can cast radioactive modelsbased on PET image segmentations, for tumours of any desired size or shape.Our results such that the 25% fixed threshold provides better accuracy forTMTV quantification of PMBCL tumours. These findings were validated throughboth the simulated and physical phantom experiments. As found in the simulation,the gradient-based algorithm overestimates tumour volume, but provides more ac-curate and consistent TLG values. However, the physical phantom experimentidentified cases where the gradient algorithm neglects significant regions of thetumour, which prompts further investigation. In a clinical setting, the 25% fixedthreshold segmentation method is recommended for routine TMTV reporting ofPMBCL.48Bibliography[1] Arman Chandra, Ramesh, Rahmim. Nuclear Medicine Physics: The Basics.Wolters Kluwer Health, 2017. → pages vii, 7, 9[2] Carlos Uribe. PHYS 404: Nuclear Medicine Lecture, 2019. → pages vii, 9[3] Shailja C. Shah, Violet Kayamba, Richard M. Peek, and DouglasHeimburger. Cancer control in low- And middle-income countries: Is it timeto consider screening? Journal of Global Oncology, 2019(5):1–8, 2019. →pages 1[4] Canadian Cancer Society. Cancer statistics at a glance, 2020. → pages 1[5] National Institute of Health. NCI Dictionaries - Cancer, 2020. → pages 1[6] Max Roser and Hannah Ritchie. Cancer, 2015. → pages 1[7] National Institute of Health. Cancer Screening, 2021. → pages 1[8] Prashanth Rawla. Epidemiology of Prostate Cancer. World Journal ofOncology, 10(2):63–89, 2019. → pages 2[9] Etienne Rousseau, Don Wilson, Fre´de´ric Lacroix-Poisson, Andra Krauze,Kim Chi, Martin Gleave, Michael McKenzie, Scott Tyldesley, S. LarryGoldenberg, and Francxois Be´nard. A prospective study on 18F-DCFPYLPSMA PET/CT imaging in biochemical recurrence of prostate cancer.Journal of Nuclear Medicine, 60(11):1587–1593, 2019. → pages 2[10] Steven P. Rowe, Katarzyna J. Macura, Esther Mena, Amanda L. Blackford,Rosa Nadal, Emmanuel S. Antonarakis, Mario Eisenberger, MichaelCarducci, Hong Fan, Robert F. Dannals, Ying Chen, Ronnie C. Mease, ZsoltSzabo, Martin G. Pomper, and Steve Y. Cho. PSMA-Based [18F]DCFPyLPET/CT Is Superior to Conventional Imaging for Lesion Detection inPatients with Metastatic Prostate Cancer. Molecular Imaging and Biology,18(3):411–419, 2016. → pages 249[11] Timothy G. Turkington, Timothy R. DeGrado, and William H. Sampson.Small spheres for lesion detection phantoms. IEEE Nuclear ScienceSymposium and Medical Imaging Conference, 4:2234–2237, 2002. → pages2, 3, 42[12] F. Hofheinz, S. Dittrich, C. Po¨tzsch, and J. Van Den Hoff. Effects of coldsphere walls in PET phantom measurements on the volume reproducingthreshold. Physics in Medicine and Biology, 55(4):1099–1113, 2010. →pages 2, 42[13] Frank P. Difilippo, Sven L. Gallo, Ryan S. Klatte, and Sagar Patel. A fillablemicro-hollow sphere lesion detection phantom using superposition. Physicsin Medicine and Biology, 55(18):5363–5381, 2010. → pages 3[14] Bert Ram Sah, Paul Stolzmann, Gaspar Delso, Scott D. Wollenweber, MartinHu¨llner, Yahya A. Hakami, Marcelo A. Queiroz, Felipe De Galiza Barbosa,Gustav K. Von Schulthess, Carsten Pietsch, and Patrick Veit-Haibach.Clinical evaluation of a block sequential regularized expectationmaximization reconstruction algorithm in 18F-FDG PET/CT studies.Nuclear Medicine Communications, 38(1):57–66, 2017. → pages 3, 11[15] Anna Dabrowska-Iwanicka and Jan A. Walewski. Primary MediastinalLarge B-cell Lymphoma. Current Hematologic Malignancy Reports,9(3):273–283, 2014. → pages 3[16] Murali Kesavan, Toby A. Eyre, and Graham P. Collins. Front-LineTreatment of High Grade B Cell Non-Hodgkin Lymphoma. CurrentHematologic Malignancy Reports, 14(4):207–218, 2019. → pages 3[17] Andrea Gallamini, Colette Zwarthoed, and Anna Borra. Positron emissiontomography (PET) in oncology. Cancers, 6(4):1821–1889, 2014. → pages 5[18] Dr. Rene Pfleger Dr. Henry Knipe. Deauville five-point scale, 2020. →pages 5[19] Jasmin Mettler, Horst Mu¨ller, Conrad Amadeus Voltin, Christian Baues,Bernd Klaeser, Alden Moccia, Peter Borchmann, Andreas Engert, GeorgKuhnert, Alexander Drzezga, Markus Dietlein, and Carsten Kobe.Metabolic tumor volume for response prediction in advanced-stage hodgkinlymphoma. Journal of Nuclear Medicine, 60(2):207–211, 2019. → pages 5[20] Qaid Ahmed Shagera, Gi Jeong Cheon, Youngil Koh, Min Young Yoo,Keon Wook Kang, Dong Soo Lee, E. Edmund Kim, Sung Soo Yoon, and50June Key Chung. Prognostic value of metabolic tumour volume on baseline18 F-FDG PET/CT in addition to NCCN-IPI in patients with diffuse largeB-cell lymphoma: further stratification of the group with a high-riskNCCN-IPI. European Journal of Nuclear Medicine and Molecular Imaging,pages 1417–1427, 2019. → pages[21] Lale Kostakoglu and Ste´phane Chauvie. Metabolic Tumor Volume Metricsin Lymphoma. Seminars in Nuclear Medicine, 48(1):50–66, 2018. → pages12[22] Baoping Guo, Xiaohong Tan, Qing Ke, and Hong Cen. Prognostic value ofbaseline metabolic tumor volume and total lesion glycolysis in patients withlymphoma: A meta-analysis. PLoS ONE, 14(1):1–17, 2019. → pages 5[23] Matthew Walker. Quantitative Dynamic 3D PET Scanning of the Body andBrain using LSO Tomographs. PhD thesis, University of Manchester, 2009.→ pages 7, 11[24] Roberto Fedrigo, Dan Kadrmas, Patricia Edem, Lauren Fougner, IvanKlyuzhin, Peter Petric, Francois Benard, Arman Rahmim, and Carlos Uribe.Optimization of Quantitative 18F-DCFPyLPET using a RealisticAnthropomorphic Phantom with Shell-less Radioactive Epoxy Lesions. InEANM Annual Meeting, Vienna, 2020. European Association of NuclearMedicine. → pages 8, 12, 17[25] Dan J. Kadrmas, Michael E. Casey, Maurizio Conti, Bjoern W. Jakoby,Cristina Lois, and David W. Townsend. Impact of time-of-flight on PETtumor detection. Journal of Nuclear Medicine, 50(8):1315–1323, 2009. →pages[26] Dan J. Kadrmas and Paul E. Christian. Comparative evaluation of lesiondetectability for 6 PET imaging platforms using a highly reproduciblewhole-body phantom with 22Na lesions and localization ROC analysis.Journal of Nuclear Medicine, 43(11):1545–1554, 2002. → pages 8, 17[27] A. Michael Morey and Dan J. Kadrmas. Effect of varying number of OSEMsubsets on PET lesion detectability. Journal of Nuclear MedicineTechnology, 41(4):268–273, 2013. → pages 8, 17[28] W. P. Segars, M. Mahesh, T. J. Beck, E. C. Frey, and B. M.W. Tsui. RealisticCT simulation using the 4D XCAT phantom. Medical Physics,35(8):3800–3808, 2008. → pages 1051[29] W. P. Segars, G. Sturgeon, S. Mendonca, Jason Grimes, and B. M.W. Tsui.4D XCAT phantom for multimodality imaging research. Medical Physics,37(9):4902–4915, 2010. → pages 10[30] Steve Ross. Q.Clear. GE Healthcare, White Paper, pages 1–9, 2014. →pages 11[31] Ronald Boellaard, Roberto Delgado-Bolton, Wim J.G. Oyen, FrancescoGiammarile, Klaus Tatsch, Wolfgang Eschner, Fred J. Verzijlbergen, Sally F.Barrington, Lucy C. Pike, Wolfgang A. Weber, Sigrid Stroobants,Dominique Delbeke, Kevin J. Donohoe, Scott Holbrook, Michael M.Graham, Giorgio Testanera, Otto S. Hoekstra, Josee Zijlstra, Eric Visser,Corneline J. Hoekstra, Jan Pruim, Antoon Willemsen, Bertjan Arends, Jo¨rgKotzerke, Andreas Bockisch, Thomas Beyer, Arturo Chiti, and Bernd J.Krause. FDG PET/CT: EANM procedure guidelines for tumour imaging:version 2.0. European Journal of Nuclear Medicine and Molecular Imaging,42(2):328–354, 2015. → pages 12[32] Go Akamatsu, Yasuhiko Ikari, Hiroyuki Nishida, Tomoyuki Nishio, AkihitoOhnishi, Akira Maebatake, Masayuki Sasaki, and Michio Senda. Influenceof statistical fluctuation on reproducibility and accuracy of SUVmax andSUVpeak: A phantom study. Journal of Nuclear Medicine Technology,43(3):222–226, 2015. → pages 14, 38[33] L. Jodal, C. Le Loirec, and C. Champion. Positron range in PET imaging:An alternative approach for assessing and correcting the blurring. Physics inMedicine and Biology, 57(12):3931–3943, 2012. → pages 15[34] A. Michael Morey, Frederic Noo, and Dan J. Kadrmas. Effect of Using 2mm Voxels on Observer Performance for PET Lesion Detection. IEEETransactions on Nuclear Science, 63(3):1359–1366, 2016. → pages 17[35] Saeed Ashrafinia, Hassan Mohy-Ud-Din, Nicolas A. Karakatsanis,Abhinav K. Jha, Michael E. Casey, Dan J. Kadrmas, and Arman Rahmim.Generalized PSF modeling for optimized quantitation in PET imaging.Physics in Medicine and Biology, 62(12):5149–5179, 2017. → pages 23[36] Andres Kaalep, Terez Sera, Sjoerd Rijnsdorp, Maqsood Yaqub, AnneTalsma, Martin A. Lodge, and Ronald Boellaard. Feasibility of state of theart PET/CT systems performance harmonisation. European Journal ofNuclear Medicine and Molecular Imaging, 45(8):1344–1361, 2018. →pages 3852[37] Matt Vanderhoek, Scott B. Perlman, and Robert Jeraj. Impact of thedefinition of peak standardized uptake value on quantification of treatmentresponse. Journal of Nuclear Medicine, 53(1):4–11, 2012. → pages 39[38] Jonathan I. Gear, Craig Cummings, Jim Sullivan, Natalie Cooper-Rayner,Patrick Downs, and Iain Murray. Radioactive 3D printing for the productionof molecular imaging phantoms. Physics in Medicine and Biology, 65(17),2020. → pages 4553Appendix ASupporting MaterialsA.1 C3PO Full Study ProtocolEquipment list• C3PO phantom + lid + screws• Cylinders (3mm, 7mm, 13mm, 17mm, 25mm, 37mm)• Food dye, funnel, liver containerProcedure1. Inject 11 MBq [18F]FDG into bladder-simulating vial in phantom.2. Close/seal phantom lid (large plastic screws should be left open).3. Fill “liver” container up to line with water (2L). Apply 3 drops of food dye.4. Inject 9 MBq [18F]FDG into container. Close lid and shake for 20 seconds.5. Use funnel to pour [18F]FDG into main phantom compartment.6. Once filling is complete, seal phantom.7. Inject 40 MBq into 250mL saline bag. Let sit for 5min to diffuse.548. Draw activity from saline bag and inject into fillable acrylic spheres.9. Insert acrylic spheres into phantom.10. Perform CTAC.11. Perform 3x3min PET frame acquisitions with time-of-flight.User Entries:Activity injected into vial (bladder)Pre-injection activity: . Time:Post-injection activity: . Time:Activity injected into 2L container (background)Pre-injection activity: . Time:Post-injection activity: . Time:Activity injected into saline bag (cylinders)Water volume removed from saline bag: .Pre-injection activity: . Time:Post-injection activity: . Time:55Figure A.1: Schematic of Canadian PET Prostate Phantom for Oncology.Phantom visualized in 3D.56Figure A.2: Schematic of Canadian PET Prostate Phantom for Oncology.Size and shape of each region is visualized.57Figure A.3: Schematic of Canadian PET Prostate Phantom. Design of eachacrylic cylinder is visualized.58


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items