Development of a simulation platform for the evaluation of PET neuroimaging protocols in epilepsy

INTERNATIONAL DOCTORAL SCHOOL OF THE USC José Paredes Pacheco PhD Thesis Development of a simulation platform for the evaluation of PET neuroimaging protocols in epilepsy Santiago de Compostela, 2022 Doctoral Programme in Neuroscience and Clinical Psychology TESIS DE DOCTORADO DEVELOPMENT OF A SIMULATION PLATFORM FOR THE EVALUATION OF PET NEUROIMAGING PROTOCOLS IN EPILEPSY José Paredes Pacheco ESCUELA DE DOCTORADO INTERNACIONAL DE LA UNIVERSIDAD DE SANTIAGO DE COMPOSTELA PROGRAMA DE DOCTORADO EN NEUROCIENCIA Y PSICOLOGÍA CLÍNICA SANTIAGO DE COMPOSTELA / LUGO 2022 DECLARACIÓN DEL AUTOR/A DE LA TESIS D./Dña. José Paredes Pacheco Título da tese: Development of a simulation platform for the evaluation of PET neuroimaging protocols in epilepsy Presento mi tesis, siguiendo el procedimiento adecuado al Reglamento y declaro que: 1) La tesis abarca los resultados de la elaboración de mi trabajo. 2) De ser el caso, en la tesis se hace referencia a las colaboraciones que tuvo este trabajo. 3) Confirmo que la tesis no incurre en ningún tipo de plagio de otros autores ni de trabajos presentados por mí para la obtención de otros títulos. 4) La tesis es la versión definitiva presentada para su defensa y coincide la versión impresa con la presentada en formato electrónico. Y me comprometo a presentar el Compromiso Documental de Supervisión en el caso que el original no esté depositado en la Escuela. En Santiago de Compostela, 20 de junio de 2022. Firma electrónica AUTORIZACIÓN DEL DIRECTOR / TUTOR DE LA TESIS Development of a simulation platform for the evaluation of PET neuroimaging protocols in epilepsy D./Dª. Pablo Aguiar Fernández .............................................................................................................. D./Dª. Núria Roé Vellvé .......................................................................................................................... INFORMA/N: Que la presente tesis, se corresponde con el trabajo realizado por D/Dª. José Paredes Pacheco, bajo mi dirección/tutorización, y a utorizo su presentación , considerando que reúne l os r equisitos exigidos en el R eglamento de Estudios de Doctorado de la USC, y que como director de esta no incurre en las causas de abstención establecidas en la Ley 40/2015. De acuerdo con lo indicado en el Reglamento de Estudios de Doctorado, declara también que la presente tesis doctoral es idónea para ser defendida en base a la modalidad de Monográfica con reproducción de publicaciones, en los que la participación del doctorando/a fue decisiva para su elaboración y las publicaciones se ajustan al Plan de Investigación. En Santiago de Compostela, 20 de junio de 2022 AGRADECIMIENTOS simulados de PET cerebral mediante técnicas de simulación de MC. En particular, o primeiro obxectivo foi validar os modelos de simulación necesarios para a implantación de tres escáneres PET comerciais. A continuación, o segundo obxectivo foi deseñar e desenvolver a propia plataforma web, integrando os modelos de escáner validados na primeira fase nunha interface gráfica con múltiples opcións para a realización dos estudos de simulación. O terceiro obxectivo foi xerar tres bases de datos simuladas de controis sans, unha para cada un dos escáneres implementados na primera parte da tese, co fin de validar o rendemento da plataforma e poñer a disposición do público en xeral estas bases de datos de imaxes PET libres. Finalmente, o cuarto obxectivo foi utilizar a plataforma para realizar un traballo de aplicación baseado na simulación dunha base de datos de estudos PET-FDG con epilepsia cortical para a súa utilización na avaliación de protocolos de PET. A primeira parte desta tese consistiu na validación dos modelos de simulación MC de diferentes escáneres PET comerciais (o GE Discovery ST, o GE Advance NXi e o Siemens Biograph mCT). Co fin de garantir que as imaxes PET simuladas fosen o máis semellantes posible ás adquiridas nun escáner PET real, as características técnicas de cada escáner real configuráronse no seu modelo simulado correspondente para posteriormente caracterizar cada escáner real e modelo simulado mediante National Electrical Manufacturers Association (NEMA). Estas probas teñen como obxectivo avaliar que o rendemento dun escáner PET se corresponde co indicado polo fabricante, establecendo os parámetros axeitados para cada adquisición. En concreto, as probas NEMA utilizadas neste traballo para as diferentes caracterizacións foron: unha proba de resolución espacial, unha proba de sensibilidade e unha proba de calidade de imaxe. Así, no caso do escáner GE Discovery ST, puidemos validar o seu modelo de simulación MC, obtendo diferenzas aceptables entre as probas NEMA reais e as probas NEMA simuladas. Estas diferenzas débense principalmente ás limitacións actuais que temos no modelo que describe a xeometría do bloque detector empregado no software de simulación e ao uso dun método de reconstrución similar ao código de reconstrucción privado (ao que non temos acceso) integrado no escáner PET real, que obviamente non é o mesmo. En canto á validación do modelo de simulación MC do escáner GE Advance NXi, esta xa foi realizada en varios traballos anteriores dentro do noso grupo de investigación, polo que adaptamos directamente estas configuracións validadas á configuración da simulación e reconstrución utilizada. No que respecta ao escáner mCT de Siemens Biograph, a validación do seu modelo de simulación MC realizouse a través de diferentes probas internas realizadas por colaboradores do Hospital Clínic de Barcelona e traballos previos doutros autores, permitíndonos integrar directamente os parámetros validados no software de simulación e de reconstrución tomográfica. O segundo e principal traballo desta tese doutoral foi o deseño e desenvolvemento da plataforma web para a simulación MC de imaxes PET: SimPET (www.sim-pet.org). Esta aplicación consiste nunha plataforma web gratuíta, accesible, eficiente, e dedicada á xeración de maniquíes dixitais e imaxes simuladas de PET, especialmente orientadas ás imaxes PET do cerebro. Para o seu desenvolvemento utilizáronse varias linguaxes de programación, como HTML5, CSS, Javascript e PHP para o front-end, MATLAB, Bash, Python e C para o backend e Laravel como framework; ademais de bibliotecas como Brain-VISET para a xeración de mapas de actividade e atenuación, SimSET para a simulación de sinogramas a partir de maniquíes de actividade e atenuación, STIR para a reconstrución das imaxes PET simuladas e SPM, FSL e NiBabel para o pre e posprocesamento das diferentes imaxes coas que se traballa. A plataforma ofrece unha interface gráfica de usuario (GUI) que permite a xeración de mapas de actividade e atenuación, sinogramas e as correspondentes imaxes PET simuladas mediante calquera dos modelos de simulación validados previamente. Ademais, conta con outras utilidades coma un visor en liña para previsualizar as imaxes xeradas na plataforma e un apartado de descrición da análise realizada. Por outra banda, o código fonte utilizado para a simulación de MC e a reconstrución tomográfica está dispoñible gratuitamente en Github e é modificable (https://github.com/txusser/brainviset_simset), o que permite que diferentes usuarios ou desenvolvedores propoñan novos modelos de simulación e de reconstrucción tomográfica, así como novas funcións que melloren as funcións actuais da plataforma. Unha vez que tivemos a plataforma implementada e a pleno funcionamento, o terceiro traballo consistiu en simular tres bases de datos de controis sans nos diferentes escáneres dispoñibles na plataforma SimPET para posteriormente validalo mediante diferentes análises estatísticas de comparación entre datos simulados e datos reais. Para xerar estudos de PET simulados o máis realistas posible, consideráronse como imaxes de entrada os mapas de actividade e os mapas de atenuación xerados co modelo GE Discovery ST, e os parámetros de entrada para a simulación (dose inxectada ao paciente, duración do estudo PET, nivel de ruído, etc.) os mesmos valores que os utilizados nos protocolos habituais de adquisición de cada un dos escáneres PET reais nos seus respectivos centros médicos. Deste xeito, xerouse unha base de datos de 25 imaxes PET do cerebro simuladas co GE Discovery ST, outra base de datos de 25 imaxes PET simuladas co GE Advance NXi e outra base de datos de 25 imaxes PET simuladas co Siemens Biograph mCT. Para validar o rendemento da plataforma, estudáronse as diferenzas entre as bases de datos simuladas e as imaxes reais de cada escáner mediante o método Bland-Altman. Ademais, tamén se realizou análise a nivel de voxel con SPM para observar diferenzas rexionais entre os grupos de imaxes reais e simulados. Como resultado de ambas análises estatísticas, obtivéronse diferenzas moi pequenas para o escáner GE Discovery ST (o 98 % dos vóxeles mostran diferenzas por debaixo do 10 %) e diferenzas relativamente pequenas para os outros dous escáneres (o 94-95 % dos vóxeles con diferenzas de menos do 10%), o que nos permitiu considerar a plataforma validada como ferramenta para xerar imaxes PET realistas do cerebro nos tres escáneres comerciais implantados. Coa plataforma SimPET xa validada con datos de controis, o cuarto e último traballo desta tese de doutoramento consistiu en simular unha base de datos de diferentes casos de pacientes que mostran regións de hipometabolismo cortical para a avaliación de protocolos en epilepsia. A metodoloxía empregada para a creación de cada un destes estudos simulados consistiu no seguinte. En primeiro lugar, seleccionouse aleatoriamente unha rexión de interese (ROI) dun atlas cerebral que contén únicamente a corteza cerebral. Unha vez seleccionada a ROI da materia gris na que queriamos inducir unha zona de hipometabolismo, consideráronse dous tipos de hipometabolismo: un deles máis focal e reducido con tres posibles niveis de afectación (60, 75 ou 90%) e outro deles máis extenso que compromete rexións con dous posibles niveis de afección (30 ou 40%). Segundo os casos, a ROI enmascarouse cun ou outro tipo de hipometabolismo e, unha vez que tivemos a ROI hipometabólica, aplicouse aos mapas de actividade realistas que xa tiñamos dispoñibles no traballo anterior para obter mapas de actividade con diferentes hipometabolismos inducidos. Estes mapas de actividade, e os correspondentes mapas de atenuación, consideráronse como entrada na plataforma SimPET para xerar estudos simulados de PET-FDG con diferentes niveis de hipometabolismo inducido utilizando o modelo GE Discovery ST. Para ter a base de datos o máis equilibrada e realista posible, inicialmente xerouse o mesmo número de estudos PET simulados para cada nivel de hipometabolismo considerado. Non obstante, tras varias revisións visuais por parte dun especialista en medicina nuclear, conseguimos afinar os hipometabolismos inducidos e obter unha base de datos formada por: 25 estudos PET cun 60% de hipometabolismo focal inducido, 31 estudos PET cun 75% de hipometabolismo focal, 19 estudos. cun hipometabolismo focal do 90%, 10 estudos PET cun hipometabolismo extenso do 30% e 15 estudos PET cun hipometabolismo extenso do 40%. A estes estudos engadimos os 25 controis simulados previamente co GE Discovery ST para ter unha base de datos total de 125 estudos de PET simulados, sendo 100 deles casos con hipometabolismo cortical simulado. Para validar esta base de datos de epilepsia cortical, realizamos unha análise cuantitativa a nivel de vóxel con SPM comparando o grupo de controis simulados fronte a cada un dos estudos de hipometabolismo simulado. Como resultado, puidemos verificar a gran semellanza entre os clusters hipometabólicos detectados por SPM nestas comparacións e a ROI hipometabólica creada para inducir hipometabolismo nos mapas de actividade orixinais. Deste xeito, non só validamos a base de datos simulada de PET-FDG na epilepsia cortical, senón que esa base de datos constitúe unha verdadeira referencia para todo tipo de traballos de avaliación de protocolos de neuroimaxe en epilepsia, xa que está formada por estudos con hipometabolismo inducido coñecido e controlado. Actualmente, esta base de datos estase a utilizar nun estudo de detectabilidade internacional realizado entre colaboradores de varios países e que ten como obxectivo avaliar a capacidade visual dos especialistas en medicina nuclear para localizar focos e zonas epileptoxénicas. Este traballo é posible grazas ao desenvolvemento dunha plataforma web desde a que colaboradores expertos poden descargar o noso conxunto de datos simulados de casos de epilepsia, realizar un informe por cada suxeito simulado indicando onde localizaron visualmente o hipometabolismo e o grao de confianza no seu diagnóstico, así como como descargar as correspondentes análises estatísticas despois do envío do informe no caso de querer comprobar a localización real e o grao de extensión do hipometabolismo inducido. Por outra banda, a plataforma SimPET tamén se utilizou para outros traballos de simulación con outros radiofármacos, como estudos de imaxe PET simulados con Florbetapir (Amiloide) ou para a simulación de bases de datos doutras enfermidades neurolóxicas tan importantes como a enfermidade de Alzheimer e a epilepsia do lóbulo temporal. Todos estes traballos demostraron a versatilidade e o potencial da plataforma SimPET como ferramenta para simular e xenerar estudos de imaxe PET, utilizando diferentes configuracións de radiofármacos e modelos de escáneres validados, e como ferramenta para avaliar protocolos de neuroimaxes PET. En definitiva, o desenvolvemento da plataforma SimPET constitúe a base dun marco no que se poden implementar futuros equipos comerciais de PET e mesmo prototipos, novos radiofármacos adicionais que xa se están a utilizar nos estudos de PET cerebral, e mesmo novas ferramentas para realizar simulacións de imaxes PET de corpo enteiro co seus radiofármacos correspondentes. Content 1 INTRODUCTION .................................................................................................................................................... 3 1.1 POSITRON EMISSION TOMOGRAPHY ....................................................................................................................... 3 1.1.1 PET Physics ............................................................................................................................................. 4 1.1.1.1 Radioisotopes and beta decay ......................................................................................................... 4 1.1.1.2 Photon-matter interaction .............................................................................................................. 4 1.1.2 PET Instrumentation ............................................................................................................................... 6 1.1.2.1 Scintillation detectors ...................................................................................................................... 7 1.1.2.2 Photomultiplier tube ....................................................................................................................... 8 1.1.2.3 Electronic collimation ...................................................................................................................... 8 1.1.2.4 Scintillator crystal performance ...................................................................................................... 9 1.1.2.5 PET scanner configurations ........................................................................................................... 10 1.1.3 PET acquisition...................................................................................................................................... 11 1.1.3.1 Line-of-response (LOR) .................................................................................................................. 11 1.1.3.2 Sinogram ........................................................................................................................................ 12 1.1.4 PET Reconstruction ............................................................................................................................... 14 1.1.4.1 Reconstruction problem ................................................................................................................ 14 1.1.4.2 Analytical reconstruction methods ............................................................................................... 15 1.1.4.3 Statistical reconstruction methods ................................................................................................ 19 1.1.5 PET corrections ..................................................................................................................................... 20 1.1.5.1 Normalization correction ............................................................................................................... 20 1.1.5.2 Attenuation correction .................................................................................................................. 21 1.1.5.3 Randoms correction ...................................................................................................................... 22 1.1.5.4 Scatter correction .......................................................................................................................... 23 1.1.5.5 Dead time correction ..................................................................................................................... 25 1.2 BRAIN IMAGING ................................................................................................................................................ 26 1.2.1 MRI ....................................................................................................................................................... 26 1.2.1.1 Clinical use ..................................................................................................................................... 26 1.2.1.2 Image analysis ............................................................................................................................... 28 1.2.2 FDG-PET ................................................................................................................................................ 31 1.2.2.1 Clinical use ..................................................................................................................................... 32 1.2.2.2 Image analysis ............................................................................................................................... 36 1.2.3 Amyloid-PET ......................................................................................................................................... 39 1.2.3.1 Clinical use ..................................................................................................................................... 40 1.2.3.2 Evaluation ...................................................................................................................................... 41 1.3 MONTE CARLO SIMULATION IN BRAIN PET ............................................................................................................ 42 1.3.1 Description ............................................................................................................................................ 42 1.3.2 Applications in brain PET ...................................................................................................................... 43 1.3.2.1 Evaluation of acquisition protocols and reconstruction methods ................................................. 43 1.3.2.2 Evaluation and optimization of quantification methods ............................................................... 44 1.3.2.3 Visual detectability studies ............................................................................................................ 45 1.3.2.4 Generation of datasets and other uses ......................................................................................... 45 1.3.3 Limitations ............................................................................................................................................ 45 1.4 JUSTIFICATION .................................................................................................................................................. 45 1.5 OBJECTIVES ...................................................................................................................................................... 47 1.5.1 Specific objectives ................................................................................................................................. 47 2 METHODS............................................................................................................................................................ 55 2.1 PATIENT DATABASES .......................................................................................................................................... 55 2.1.1 Our cohort (CIMES) ............................................................................................................................... 55 2.1.2 Our cohort (CHUS) ................................................................................................................................ 56 2.1.3 Our cohort (Hospital Clinic de Barcelona)............................................................................................. 56 2.1.4 ADNI ...................................................................................................................................................... 56 2.2 MC SIMULATION .............................................................................................................................................. 56 2.2.1 SimSET code .......................................................................................................................................... 57 2.3 TOMOGRAPHIC RECONSTRUCTION SOFTWARE ......................................................................................................... 58 2.3.1 STIR ....................................................................................................................................................... 58 2.4 CHARACTERIZATION OF PET SCANNERS .................................................................................................................. 58 2.4.1 NEMA protocols .................................................................................................................................... 59 2.5 WEB-BASED PLATFORM ...................................................................................................................................... 62 2.6 GENERATION OF HEALTHY DATABASES ................................................................................................................... 63 2.7 GENERATION OF AN EPILEPSY DATABASE ................................................................................................................ 63 2.8 IMAGE PROCESSING AND QUANTIFICATION ............................................................................................................. 66 2.8.1 SPM ....................................................................................................................................................... 66 2.9 STATISTICS ....................................................................................................................................................... 68 2.9.1 Two-sample t-test ................................................................................................................................. 70 2.9.2 Paired t-test .......................................................................................................................................... 70 2.9.3 Bland-Altman comparison analysis ....................................................................................................... 70 3 RESULTS .............................................................................................................................................................. 74 3.1 VALIDATION OF MC SIMULATION MODELS ............................................................................................................. 74 3.1.1 Validation of GE Discovery ST ............................................................................................................... 74 3.1.2 Validation of GE Advance NXi ............................................................................................................... 76 3.1.3 Validation of Siemens Biograph mCT .................................................................................................... 76 3.1.4 Validated MC simulation models .......................................................................................................... 76 3.2 SIMPET - AN OPEN ONLINE PLATFORM FOR THE MONTE CARLO SIMULATION OF REALISTIC BRAIN PET DATA...................... 79 3.2.1 SimPET platform ................................................................................................................................... 79 3.2.2 GUI and functionalities ......................................................................................................................... 79 3.2.3 Internal processing ............................................................................................................................... 81 3.3 GENERATION OF HEALTHY DATABASES ................................................................................................................... 82 3.4 GENERATION OF AN EPILEPSY DATABASE ................................................................................................................ 84 4 DISCUSSION ........................................................................................................................................................ 91 4.1 VALIDATION OF MC SIMULATION MODELS ............................................................................................................. 91 4.2 SIMPET - AN OPEN ONLINE PLATFORM FOR THE MONTE CARLO SIMULATION OF REALISTIC BRAIN PET DATA...................... 93 4.3 GENERATION OF HEALTHY DATABASES ................................................................................................................... 94 4.4 GENERATION OF AN EPILEPSY DATABASE ................................................................................................................ 95 5 CONCLUSIONS .................................................................................................................................................... 99 6 BIBLIOGRAPHY .................................................................................................................................................. 103 7 APPENDICES ...................................................................................................................................................... 129 7.1 SIMULATION OF PATHOLOGICAL MAPS ................................................................................................................ 129 7.2 AMYLOID-PET SIMULATION .............................................................................................................................. 130 8 DECLARATIONS: USE OF IMAGES, TABLES AND PUBLISHED CONTENT ............................................................ 133 8.1 JOURNAL AUTHORIZATIONS FOR THE USE OF IMAGES AND TABLES ............................................................................. 133 8.2 LIST OF THESIS PUBLICATIONS ............................................................................................................................ 155 8.2.1 SimPET – An online platform for the Monte Carlo simulation of realistic brain PET data. Validation for 18F-FDG scans ............................................................................................................................................. 155 9 CONFLICT OF INTEREST .................................................................................................................................... 159 INTRODUCTION Introduction 9 Figure 6: Diagram of a coincidence detection. Reproduced with permission of Springer Nature from (Cherry & Dahlbom, 2006). 1.1.2.4 Scintillator crystal performance There are certain detector properties that determine the performance of the PET scanner and will influence the quality of your images: a) The time resolution is the statistical uncertainty or timing error between the arrival of a photon at the scintillator and output pulse leading edge. When a photon is detected by a scintillator crystal, the system initiates a coincidence time window. If another photon is received within this time window, it will be considered that both photons come from the same annihilation. The composition of the scintillator crystal, and in particular, the decay time, is directly related the timing resolution of the detector block and thus the coincidence time window. The short decay times lead to better timing resolution and shorter coincidence time window (Bailey et al., 2005). b) The intrinsic efficiency is the ability to detect photons reaching the scintillator detector per unit of activity. It is mainly determined by the stopping power of the crystals, which is related to the composition of the scintillator crystal. c) The energy resolution is the ability of a detector to estimate the energy of an arriving photon. This is essential in PET detectors for detecting separately (and then removing) those photons that achieve the detector after Compton interactions in the tissue. High light output from scintillation crystals lead to better energy resolution. JOSÉ PAREDES PACHECO 10 d) Regarding the material used, BGO has been one of the most used due to its high stopping power, but its long decay time and low light output is a huge limitation, and nowadays LSO and GSO are usually preferred (see Table 1) (Saha, 2010a). 1.1.2.5 PET scanner configurations As we have seen, a PET scanner is made up of detector blocks grouped into arrays. The distribution of these sets of detectors determines the geometry of the PET scanner. The most common design is the ring geometry around the radioactive source. However, the first scanners were two flat detectors facing each other and later they evolved to other designs such as a rotating partial ring system, a hexagonal array of panel detectors or continuous curved panel systems (see Figure 7). Figure 7: Four different PET scanner designs: fixed ring design (a), rotating partial ring system (b), fixed system with six flat detectors (c) and fixed system using continuous curved panels. Reproduced with permission of Springer Nature from (Cherry & Dahlbom, 2006). The geometry of a PET scanner is of paramount importance as the overall detection efficiency of scanner is directly related to the distance between the source and the detectors (transverse or transaxial plane), the diameter of the ring, the number of detector blocks per ring, and the number of consecutive rings in the direction of the patient's body (axial plane). Nonetheless, the angular coverage of the current PET scanners is limited and the scanner only covers a small part of the patient's body, so they are usually equipped with beds to move the patient along the detector, so that whole-body PET studies have to be acquired by using multiple bed positions (Cherry & Dahlbom, 2006). Introduction 11 1.1.3 PET acquisition 1.1.3.1 Line-of-response (LOR) Once two photons are detected in coincidence, it can be defined the line that joins the two centers of the detectors that receive the photons. It is called line of response (LOR) and represents the line where the annihilation has occurred. Due to limitations in energy and timing resolution mentioned above, different types of LORs can be detected (Bailey, 2005): a) Singles (Figure 8a), when only one photon is detected in the coincidence time window. It occurs when the other photon coming from the same annihilation is lost (outside of the angular coverage of the detectors). It is a fairly frequent event in PET. b) True coincidences (Figure 8b), when two photons generated from the same annihilation reach two opposite detectors without matter interaction and are registered within the same coincidence time window. In this case, the LOR corresponds to the path followed by the photons and that joins the two detectors. c) Scatter coincidences (Figure 8c), when one or both photons detected within the same coincidence have been scattered by Compton effect, losing a small part of the energy. Most PET scanners do not have enough energy resolution to discard these scattered photons. These scatter coincidences produces a loss of image contrast (Humm et al., 2003). d) Random coincidences (Figure 8d), when two photons generated from different annihilations are detected within the same coincidence time window. This generates an incorrect LOR that produces a background constant signal in the final image. If 𝑎 and 𝑏 are the detectors that register both photons, the random event count rate (𝑅𝑎𝑏) can be expressed as a function of the number of photons that 𝑎 and 𝑏 can detect (𝑁𝑎 and 𝑁𝑏) and the coincidence window width (2𝜏): 𝑅𝑎𝑏 =2𝜏∗𝑁𝑎∗𝑁𝑏 (5) As 𝑅𝑎𝑏 is proportional to 2𝜏, the better timing resolution the detector has, the fewer randoms events are generated. e) Multiple coincidences (Figure 8e), when three or more photons are detected within the same coincidence time window. They are normally discarded without degradation effects on image quality, but results in loss of true coincidences. For the calculation of the prompt count rate, in addition to the true coincident events, the random and scatter events are also considered valid since only two photons are detected in the same coincident time window. Instead, single and multiple events are directly discarded from the data. The best performance is achieved when true events are recorded and used for image formation, and, therefore, various correction methods for random and scatter events have to be applied. JOSÉ PAREDES PACHECO 12 Figure 8: Types of coincident events in PET: Singles (a), trues (b), scatter (c), random (d) and multiple (e) Adapted with permission of Springer Nature from (Bailey, 2005). 1.1.3.2 Sinogram The last step in PET data acquisition is storing the accepted data. All the LORs passing through the patient are stored and defined by an orientation angle (𝜙) with respect to the axial plane and a distance to the center of the scanner (𝑠), forming a histogram of the LORs called sinogram (R. Schmitz et al., 2013). If for each point of the patient we represent the detected LORs as a function of these two parameters in a two-dimensional x-y graph, we will form the sinogram for that transaxial plane (Figure 9). Each value of the sinogram represents the number of matches detected along the LOR defined by the angle (x-axis) and the distance to the center of the scanner (y-axis). Figure 9: Conversion of detected raw data to a sinogram. Adapted with permission of JNMT from (Fahey, 2002). Introduction 13 In addition to the sinograms, the information detected in the LORs can also be saved as projections. These projections 𝑝(𝑠,𝜙) represent the parallel line integrals of the activity for each given angle 𝜙. They can be obtained by going through each row of the sinogram. The clinical scanners are made up of several consecutive rings and, therefore, LORs have to be defined based on more parameters than for a single full ring of detectors. In this case, a LOR can be defined as 𝐿(𝑠,𝜙,𝜃,𝑧) where 𝑠 and 𝜙 are the variables that we saw previously and that define the plane of the sinogram, 𝜃 is the angle of inclination with respect to the axial plane (z-axis) and 𝑧 is the position on the z-axis (see Figure 10). This allows different ways of acquiring PET images. The two-dimensional mode only accepts coincident events in direct planes (polar angle 𝜃=0), i.e. between detectors of the same ring (see Figure 11a). In this mode N sinograms are acquired, where N is the number of rings. Therefore the sensitivity of the scanner is limited to N times the sensitivity of a ring. A more standard 2D mode is the one that accepts the coincidences generated in the direct planes, but also in the crossed planes, that is, between detectors of two adjacent rings. Thus, in a scanner with N rings, we can define N direct planes plus N-1 crossed planes, being a total of 2N-1 planes of coincidences, improving the sensitivity (Cherry & Dahlbom, 2006). Initially thin tungsten shields, known as septa, were placed between the rings to reduce the detection of photons from larger angles and minimize the number of scattered and random photons in early 2D mode scanners (see Figure 11b) but currently it is not like that. Figure 10: Diagram for a multi-ring PET camera. Reproduced with permission of Springer Nature from (Bailey, 2005). JOSÉ PAREDES PACHECO 14 Increasing the number of coincidence planes significantly improves the sensitivity of the PET scanner. For this reason, today most of the PET scanners acquire in three-dimensional mode (fully 3D mode). In this mode there are no septa and coincident event planes are allowed between any combinations of rings (see Figure 11c). This leads to N2 sinograms being generated. Despite generating a lot of redundant information, this improvement in sensitivity relies on a better SNR in the image, reducing the acquisition time and the amount of injected dose (Cherry & Dahlbom, 2006). Figure 11: More common acquisition modes: Fully 2D (a), 2D (b) and 3D (c). Adapted with permission of Elsevier from (Zanzonico, 2004). 1.1.4 PET Reconstruction Tomographic reconstruction consists of recovering the 3D distribution of the radiotracer using the projections or sinograms. This section discusses the reconstruction problem and the most common types of reconstruction methods. We start with analytical reconstruction to go on to reconstruction with statistical methods. 1.1.4.1 Reconstruction problem PET reconstruction consists of solving the following linear problem: 𝑝=𝐻𝑓+ 𝑛 (6) where 𝑝 are the acquired data, 𝐻 is known as the system response matrix (SRM), 𝑓 is the unknown 3D distribution of the radiotracer and 𝑛 are the possible errors in the observation. Introduction 15 The goal is to use 𝑝 to obtain 𝑓. Although the acquired data are stochastic variables, as a first approximation, we can assume that the data are deterministic, that is, without a source of error. Therefore, in equation 6, 𝑛 would be a known number and we could find the exact solution of the equation for 𝑓. Analytical reconstruction methods use this to find a direct solution through the image 𝑓 from the projections 𝑝. The advantage of assuming deterministic data is that it greatly simplifies the reconstruction process, making it faster. The bad part is that noise information is discarded. 1.1.4.2 Analytical reconstruction methods In 2D analytical reconstruction methods the basic step is backprojection. This procedure consists of placing each value of a projection 𝑝(𝑠,𝜙) back to its corresponding LOR. However, we do not know how the values were distributed in the original LOR, so the approach we can do is to put a constant value along the entire LOR. In Figure 12 we can see an example of backprojection 𝑏(𝑥,𝑦,𝜙) for a specific angle 𝜙. Figure 12: Backprojecton of all values of a projection 𝒑(𝒔,𝝓) for a fixed value of 𝝓. Reproduced from (Alessio & Kinahan, 2005). Mathematically, for a given angle 𝜙, a projection can be defined by the Radon transform of the radiotracer distribution 𝑓(𝑥,𝑦), as follows (Kak & Slaney, 2001): 𝑝(𝑠,𝜙)=∫ ∫ 𝑓(𝑥,𝑦)∙𝛿(𝑥 𝑐𝑜𝑠𝜙+ 𝑦 𝑠𝑖𝑛𝜙− 𝑠)𝑑𝑥𝑑𝑦 ∞ −∞ ∞ −∞ (7) where 𝛿 is the Delta function, so the integral is always zero except in the LOR 𝐿(𝑠,𝜙). If we integrate along all the angles, we get the backprojected image 𝑏(𝑥,𝑦): 𝑏(𝑥,𝑦)= ∫ 𝑝(𝑠,𝜙)|𝑠=𝑥𝑐𝑜𝑠𝜙+𝑦𝑠𝑖𝑛𝜙𝑑𝜙 𝜋 0 (8) JOSÉ PAREDES PACHECO 16 The integral travels 180º since the other angles do not give new information. Figure 13: Original image VS Backprojected image. Adapted from (Zvolský, 2014). If the backprojection is calculated for all the required angles, we might think that we would recover the image with the radiotracer distribution, however, this is not the case. Basic backprojection has the problem of intensity oversampling in the center of the image and less sampling at the edges, producing a blurring effect (see Figure 13). Figure 14: Two-dimensional CST where the 1-D Fourier transform of a projection at angle 𝝓 is equivalent to the 2-D Fourier transform of the central section at the same angle of the radioactive source. Adapted with permission of Elsevier from (Kinahan et al., 2004). This blurring effect can be minimized by applying a filter to the data during acquisition. This filtered data is then backprojected to reconstruct an image more similar to the original source. This reconstruction method is called Filtered Backprojection (FBP) and it is one of the 2D reconstruction methods most used today due to its efficiency and better precision. The FBP method is based on the Fourier transform of the acquired data as well as the central slice theorem (CST). According to the CST (Kinahan et al., 2004), it is verified that: Ƒ1{𝑝(𝑠,𝜙′)}=Ƒ2{𝑓(𝑥,𝑦)}|𝜙=𝜙′ (9) Introduction 17 where Ƒ1 is the 1-D Fourier transform of the projection 𝑝(𝑠,𝜙) and Ƒ2 is the 2-D Fourier transform of the radiotracer distribution 𝑓(𝑥,𝑦). This implies that doing the 1-D Fourier transform of the projection 𝑝 for an angle 𝜙′ is equivalent to doing the 2-D Fourier transform of the data 𝑓(𝑥,𝑦) and taking a slice through the origin with the same angle 𝜙′. A schematic graph of this can be seen in Figure 14. Following the CST, using the inverse Fourier transform and switching to polar coordinates, the real distribution, 𝑓(𝑥,𝑦), can be expressed as: 𝑓(𝑥,𝑦)=Ƒ2 −1{𝐹(𝑣𝑥,𝑣𝑦)}= ∫ 𝑑𝜙 𝜋 0 [∫ 𝑑𝑤|𝑤|𝑃(𝑤)𝑒2𝜋𝑖𝑤𝑠 ∞ −∞ ]=∫ 𝑑𝜙𝑝′(𝑠,𝜙) 𝜋 0 (10) where we 𝐹(𝑣𝑥,𝑣𝑦) is the Fourier transform of 𝑓(𝑥,𝑦), 𝑃 is the Fourier transform of 𝑝(𝑠,𝜙) and 𝑝′(𝑠,𝜙) is the original projection but multiplied by a ramp frecuency filter |𝑤| in the Fourier space (Zvolský et al., 2017). In this way, the FBP method consists of applying the Fourier transform to the projection 𝑝(𝑠,𝜙), applying the ramp filter |𝑤|, coming back to the initial space with the inverse of the Fourier transform, this filtered projection is backprojected and then all the backprojected projections are added to generate the FBP image (see Figure 15). Figure 15: Filtered backprojection method. Reproduced from (Zvolský, 2014). JOSÉ PAREDES PACHECO 18 The FBP method is widely used for 2D reconstruction thanks to its simplicity and easy implementation. However, operating in 2D mode is not the most appropriate, since working in 3D mode takes more advantage of the radiation emitted by the patient, increasing the sensitivity of the scanner. For this reason, various reconstruction methods take advantage of data acquired in 3D and putting them together in parallel transverse sinograms (𝜃=0) so that they can be reconstructed with the traditional FBP method. These algorithms are known as rebinning algorithms. The simplest of these algorithms is the single-slice rebinning (SSRB), which takes the mean position where an event has occurred in the axial plane and adds it to the corresponding direct sinogram closest to that axial position. However, this algorithm only works well when the radioactive source is small, but in the case of patients some blurring of the data occurs in the axial direction (Cherry & Dahlbom, 2006). Another method is multi-slice rebinning (MSRB), where the LORs are rebinned at different direct sinograms, increasing the resolution but being too sensitive to noise artifacts (Biersack & Freeman, 2008). The most widely accepted rebinning algorithm is the Fourier rebinning (FORE). This algorithm is based on relating the 2D Fourier transform of the oblique sinograms (𝜃>0) with the 2D Fourier transform of the transverse sinograms (𝜃=0). In particular, Fourier transformed oblique sinograms can be rebinned into direct sinograms and, after sample normalization in the Fourier space, the inverse transform can be applied to retrieve direct sinograms (Alessio & Kinahan, 2005). FORE was the rebinning method implemented in most scanners instead of SSRB and MSRB, since it supports greater angles and is more accurate in larger radioactive sources. Another option for analytical 3D image reconstruction is the 3-D reprojection algorithm (3DRP). It is considered as an extension of the FBP method. The difference between 2D and 3D acquisition is that, while in 2D you have the information of all the projections for any angle (𝜙 ranging from 0 to 180º), in 3D mode different projections are truncated during the acquisition when 𝜃>0 (see Figure 16). This is due to the cylindrical geometry of most PET scanners. The solution is the data redundancy presents in 3D acquisition. Incorporating oblique sinograms improves the SNR of the reconstructed image and, in this case, allows to solve the problem of truncated projections. In the 3DRP method the 2D sinograms (𝜃=0) are extracted and reconstructed using FBP to get a 3D image volume. This image can be projected to add activity in oblique LORs lost during acquisition and remove truncation. This is called reprojection or forward-projection. Now the projections are 2D and represent the activity for a particular 𝜙 and 𝜃. The 2D Fourier transform is applied to these projections. Then, they are multiplied by a 2D filter and return to the original space taking inverse 2D Fourier transform to finally be backprojected into a 3D image matrix. The final reconstructed 3D image is obtained repeating this process for each supported 𝜙 and 𝜃 (Cherry & Dahlbom, 2006). Figure 16: Truncation of projections in axial plane when 𝜽>𝟎. Modified with permission of Springer Nature from (Cherry & Dahlbom, 2006). Introduction 25 d) MC Simulation methods. A better way to use simulation to correct scatter events is through Monte Carlo (MC) simulations. Unlike previous analytical simulation methods, MC simulations generates scattered and unscattered data separately. This makes it a great tool to perform scatter correction for any specified attenuation and emission distribution or PET configurations. In this method, a 2D reconstruction of the blank and transmission data or a CT image is used as the attenuation map and an image reconstructed by FBP as an initial estimate of the activity distribution. This initial image is assumed to represent the true radiotracer distribution with no scatter events. The algorithm then generates photon pairs according to the initial activity distribution and tracks the random interactions through the scattering medium according to the attenuation map. From the simulation, the scattered sinograms are generated, which are smoothed and removed from the measured data to obtain the unscattered data. This process can be repeated to reduce the initial error by assuming that the reconstructed image has no scatter. MC is one of the most accurate that we have discussed. However, it requires great computational power to carry out the necessary simulations and to have sufficient counting statistics for an accurate scatter correction, since a large part of the tracked photons will undergo energy or trajectory changes that will impair their detection (Meikle & Badawi, 2005; Saha, 2010a). 1.1.5.5 Dead time correction Between the detection of one coincidence event and the next one, the PET scanner and in particular the detector block need a recovery time. At this time that goes from a photon interacts with the scintillator crystal until it is processed as a count, it is classified as a coincidence event when another photon reaches another detector in the same time window and the detector is again available to receive the next event is called dead time of the PET scanner. Dead time sources are highly dependent on the architecture and design of the PET scanner. An influencing factor is the integration time of the charge of a photon in the PMT after depositing the energy on the scintillator crystal. If a new photon reaches the crystal while the previous one is still being integrated (called pulse pileup phenomenon) it can happen that both events are discarded or considered as a single event but with incorrect energy and position (Figure 19), which leads to distortions in the image. There are other factors such as the “reset” time of the electronics detectors or the time to process a coincident event that do not allow accepting new coincidences and contribute significantly to the scanner dead time. To improve in this sense, pulse pileup rejection circuits, high-speed electronics or buffers to retain overlapping events during the dead time are often used (Saha, 2010a). Figure 19: Pulse pileup phenomenon. Adapted with permission of Springer Nature from (Meikle & Badawi, 2005). JOSÉ PAREDES PACHECO 26 A basic option to correct the dead time is to collect data from observed count-rates of decay sources, calculate correction factors and apply them to the acquired data. However, this method does not take into account the spatial variations in the distribution of activity that can alter the count-rate in each part of the system. Therefore, a more precise option is to construct a model where the correction factors from all sources are applied to each subsystem of the scanner (Meikle & Badawi, 2005). Dead time models are often made up of two components: paralyzable and non-paralyzable (Knoll, 2000). The paralyzable time refers to when the system cannot process any events for a time after each event, but the system is not considered dead. For example, when the charge of a photon is being integrated into a PMT but more photons arrive to deposit energy on the crystal, which must decay before being able to process the next event. In the non-paralyzed case, the system is considered dead, so additional events are ignored. These two components are often used in series in PET scanners. 1.2 BRAIN IMAGING Since this thesis is focused on epilepsy, in this section we discuss some of the main types of brain images that have been gradually introduced into the clinical routine for the study of epilepsy and other important neurological diseases and how they complement each other to achieve a better diagnosis. We will also see different ways of evaluating these studies to observe the changes caused in the brain due to the disease, be able to differentiate pathologies or even obtain information on its initial states. 1.2.1 MRI MRI has become an essential tool for diagnosis and research in medicine, becoming the gold standard for the study of brain structures, with a particular focus on the study and differentiation of neurological and psychiatric diseases such as epilepsy, dementias, parkinsonisms and other disorders. The main advantage of MRI compared to other types of images is its great spatial and temporal resolution, showing a strong contrast between the different tissues of the brain, something impossible in other modalities. 1.2.1.1 Clinical use a) Epilepsy In clinical routine, structural MRI is an essential tool for the diagnosis and study of the underlying pathophysiology of epilepsy, since it contributes greatly to the detection of the structural brain abnormalities involved. In the most common type of epilepsy, temporal lobe epilepsy (TLE), MRI favors the detection of gliosis and impairments in the temporal lobe, hippocampus and amygdala (R. Kuzniecky et al., 1987; D. H. Lee et al., 1998). In frontal lobe epilepsy, focal frontal lobe or multilobar abnormalities usually appear (Lorenzo et al., 1995) and in occipital lobe epilepsy, a less frequent type, occipital cortical dysplasia or occipital periventricular heterotopia among other syndromes (Taylor et al., 2003). In Figure 20 we can see some examples of the different types of epilepsy and their manifestations on MRI. Furthermore, MRI has an important role in presurgical evaluation in patients with refractory epilepsy (seizures cannot be properly controlled using antiepileptic drugs). MRI can help visualize and detect over 60% of frontal lobe lesions and over 80% of temporal lobe lesions (Álvarez-Linera Prado, 2012). In particular structural impairments such as hippocampal sclerosis, cortical dysplasias or focal lesions are good indicators on MRI to select candidates for surgery since surgical treatment of focal cortical malformations has a high probability of cure (Duncan, 1997; Dupont & Baulac, 2004; R. I. Kuzniecky, 2005). Introduction 27 Figure 20: Coronal MR images show a decrease in the volume of the left hippocampus in a case of TLE (a), a focal cortical dysplasia of the left superior frontal sulcus in a case of frontal lobe epilepsy (b) and bioccipital periventricular heterotopia in a case of occipital lobe epilepsy. Adapted with permission of Wolters Kluwer Medknow Publications and Oxford University Press from (Singh et al., 2013; Taylor et al., 2003; Winston et al., 2013). b) Other clinical uses Regarding the study of other neurological diseases such as dementia, MRI can show the degree of anatomical deterioration of different regions or reveal atrophy patterns associated with the appearance of certain dementias (Raposo Rodríguez et al., 2018). Among all dementias, Alzheimer's disease (AD) is the most widespread in the population, frequently being the phase after mild cognitive impairment (MCI). Numerous structural imaging studies have been performed to find brain atrophies and other factors that characterize the progression of MCI to AD (Frankó et al., 2013; S. H. Lee et al., 2016; Whitwell et al., 2008). In Figure 21 we can see T1-weighted MRI images of how gray matter (GM) atrophy evolves from a healthy subject to a subject with MCI and a subject with AD. Figure 21: T1-weighted images showing GM volume reduction in a healthy subject, in a MCI patient and in an AD patient. Reproduced with permission of Springer Nature from (Chandra et al., 2019). Another of the main advantages of using MRI for the study of dementias is to be able to observe the patterns of structural deterioration and discriminate between AD and other dementias as frontotemporal lobe dementia (FTLD) or dementia with Lewy bodies (DLB). In Figure 22 we can see examples of some characteristic patterns of AD, FTLD and DLB. JOSÉ PAREDES PACHECO 28 Figure 22: Atrophy patters in AD (A), DLB (B) and FTLD (C). Coronal T1-weighted images show cortical deterioration in general but with greater atrophy in the hippocampus in AD (A) than in DLB (B). Axial T1weighted image show a deterioration in the frontal and prefrontal cortices in FTLD case (C). Adapted with permission from (Ayers et al., 2019). In relation to Parkinson's disease (PD), traditional MRI does not usually show changes in the anatomy of the brain. For this reason, more advanced MRI techniques are often considered, such as functional MRI or diffusion imaging (Pyatigorskaya et al., 2014). 1.2.1.2 Image analysis a) Quantification methods In MRI, different patterns of sclerosis, lesions or dysplasias are associated with specific diseases. Traditionally, visual analysis has been the most efficient method to analyze these pathologies. However, this does not help to refine the degree of atrophy present (Vemuri & Jack, 2010) or it may not allow detect the lesion (Muhlhofer et al., 2017; Wang & Alexopoulos, 2016). It is also a subjective evaluation and time consuming used in large groups of subjects. This has led to the development of automatic and semi-automatic quantitative techniques that allow post-processing and evaluating morphological changes in single subjects with greater efficiency and precision. One of the most popular options is voxel-based morphometry (VBM) (Ashburner & Friston, 2000) using T1-weighted images because it provides biologically reliable results by performing statistical tests at voxel level in the whole brain. The most common situation is to find atrophy patterns between a patient and a group of healthy controls. Then a t-test would be performed on each voxel of the brain to obtain regions with significant morphological differences. Before the statistical analysis, it is necessary to preprocess the images, since we need the location of each anatomical structure to be the same in all subjects so that they coincide at the voxel level. The basic pipeline has the following steps: segmentation, spatial normalization and smoothing of the images (Whitwell, 2009), as can be seen in Figure 23: a. Inhomogeneity correction and segmentation: an inhomogeneity correction is applied to the T1-weighted images. The images are then segmented into different tissue compartments: GM, white matter (WM), cerebrospinal fluid and others. b. Spatial normalization: it consists of applying to each segmented image a series of mathematical transformations to place all the images of the different subjects in the same standard space. In particular, a set of linear transformations are Introduction 29 performed to transform the space followed by non-linear deformations. In this way, the images will not only be in the same standard space, but the anatomical structures will have a similar shape, so that they coincide point by point. c. Spatial smoothing: the segmented images in standard space are smoothed, so that the intensity value in each voxel is weighted by the average of the adjacent voxels, obtaining a blurred image of each segmented image. This helps to decrease intersubject variability during statistical analysis. Once the preprocessing is done, the images are ready to be used in the statistical analysis. This analysis is usually a t-test where the null hypothesis is that the difference of a tissue volume between the groups is zero, although depending on the images and parameters with which we work, there may be other types of tests. Depending on the chosen p-value, the t-test will generate a map of t-statistical values with all voxels where the null hypothesis is not fulfilled and, therefore, there are significant differences between the groups. Figure 23: Workflow of a VBM analysis using T1-weighted images. Reproduced with permission of Elsevier from (Kurth et al., 2015). JOSÉ PAREDES PACHECO 30 b) Examples In epilepsy, it has provided essential information for the detection of cortical abnormalities, especially in nonlesional MRI situations in the context of epilepsy pre-surgical evaluation (Wang & Alexopoulos, 2016). In cases of TLE, the VBM analysis usually shows an asymmetrical volume distribution mainly in hippocampus and less frequently in parahippocampal, entorhinal cortex and contralateral hippocampus (Keller & Roberts, 2008; Mueller et al., 2006) (Figure 24). In extratemporal lobe atrophies, there is a more bilateral distribution of lesions mainly affecting the parietal lobe and thalamus (Keller & Roberts, 2008). Figure 24: Differences in the volume of GM between a control group and a group with TLE. The loss of volume extends to hippocampus, extrahippocampal and extralimbic areas. Adapted from (Mueller et al., 2006). Introduction 31 Regarding dementias, the VBM analysis is a useful tool to characterize the reduction of GM volume in patients compared to a control group, across the whole brain. Atrophic changes in temporal lobe structures (hippocampus, parahippocampal cortex, lingual and fusiform gyri, amygdala) are commonly found in multiple studies for cases of MCI and AD (Busatto et al., 2008; Chételat et al., 2005; Ferreira et al., 2011; Yao et al., 2011). Another notable role of VBM analysis in dementias is its usefulness in the differentiating diagnosis between AD and other dementias such as FTLD or DLB. In FTLD, the VBM analysis shows evidence that GM patients undergo volume reduction in frontal cortex, insula, and striatum (Pan et al., 2012), less frequent areas in AD. While in DLB patients, the higher level of atrophy in the dorsal brain stem and the better preservation of the medial temporal lobe, hippocampus and amygdala compared to AD patients improve differentiation (see Figure 25) (Burton et al., 2002; Matsuda et al., 2019; R. Watson et al., 2012). Figure 25: Statistical parametric maps of the most preserved areas in a group of DLB patients compared to a group of AD patients. Adapted with permission of Elsevier from (Burton et al., 2002). The main limitation of the VBM is that it is designed to test differences between groups, being neither optimized nor sufficiently validated for the detection of areas with great variability between patients, such as in cases of epilepsy with focal cortical dysplasia (Yasuda et al., 2010), for the diagnosis of individual cases. There is validated software for performing VBM in single subject cases, such as Neurocloud Vol (https://www.qubiotech.com/soluciones/neurocloudvol/), but it is not present in all clinical centers. Despite this limitation, the VBM analysis is an important tool for the investigation of different aspects in groups of patients, providing a better knowledge of the morphological damage suffered in the brain and complementary to other modalities and types of analysis, as we will see in the following sections. 1.2.2 FDG-PET The compound 2-Deoxy-2-[18F]fluoro-D-glucose (FDG) consists of a molecule similar to glucose where an OH group has been replaced by the radioisotope 18F. FDG, like glucose, enters cells through glucose transporters. FDG and glucose are phosphorylated by hexokinase. However, unlike glucose, the FDG molecule is not metabolized and is trapped in tissue (see Figure 26). JOSÉ PAREDES PACHECO 32 Figure 26: Comparison of glucose and FDG cellular uptake and metabolism. Reproduced with permission of Elsevier from (Filho et al., 2011). As glucose is the main source of energy for the brain, this trapping makes possible to study cellular metabolism and see the physiological and biochemical changes in the brain. The alteration of brain metabolism in different areas would reflect an alteration in synaptic activity resulting from a combination of processes involved in the pathogenesis of neurodegenerative diseases. For this reason, FDG is one of the most widely used biomarkers for diagnosis in neurology. 1.2.2.1 Clinical use The main indications for the use of the brain FDG-PET are presurgical evaluation of refractory epilepsy, early and differential diagnosis of dementia and differential diagnosis between Parkinson disease (PD) and other atypical parkinsonisms; less common cases of patients with stroke or psychiatric disorders (Schöll et al., 2014). a) Epilepsy FDG-PET has proven to be a valuable tool in clinical routine to support MRI in the non-invasive localization of epileptogenic brain regions in intractable epilepsy for surgery for many years (Casse et al., 2002; Willmann et al., 2007), especially in cases where no signs of lesions on MRI (Cendes, 2013; Jaisani et al., 2020; Muhlhofer et al., 2017; Rathore et al., 2014) proving to have a higher sensitivity than other kind of images (Sarikaya, 2015). In particular, Introduction 33 the documented sensitivity of FDG-PET is around 85% -90% in cases of TLE, while it drops to values around 55% in frontal lobe epilepsy or even lower in cases of occipital lobe epilepsy (Kumar & Chugani, 2017). In any case, it is a very useful imaging modality for presurgical evaluation, especially if it is supported by quantitative analyzes to improve diagnostic confidence. Sometimes it is the only imaging method with a positive result to evaluate the option of surgical intervention. Figure 27: Interictal FDG-PET study in a patient with temporal lobe epilepsy. Reproduced with permission of Elsevier from (Kumar et al., 2012). The most common is to perform the acquisition of the FDG-PET study in an interictal period, due to the short life of FDG and technical difficulties in acquiring ictal studies in patients with infrequent epileptic seizures. The characteristic observed in the FDG-PET studies are small areas or foci with low glucose uptake (hypometabolism) during the interictal period (Figure 27). The reasons why these hypometabolic foci appear during the interictal period are not clear, but they may be due to different mechanisms such as neuronal or synaptic density loss (Willmann et al., 2007). Ictal PET, although less frequent, can also be performed. Several authors (Meltzer et al., 2000; Nooraine et al., 2013) have reported ictal PET studies where areas of hypermetabolism have been observed in frontal and occipital regions. The correlation between these hypermetabolic areas and the good postoperative results support the usefulness of ictal PET, especially in cases of frontal or occipital epilepsy (Ladrón De Guevara, 2013). Some limitations of FDG PET for the localization of epileptogenic foci are: the low resolution of PET studies and the fact that the observed changes in metabolism usually extend beyond the real ictal zone. Therefore, it is important to rely on other imaging modalities when defining surgical borders. In addition to the location of epileptogenic foci, FDG-PET also shows the general status of the other areas of the brain. This may lead to correlations between epileptogenic zones and the malfunction of other regions of the brain or functional deficit zones (Jaisani et al., 2020; Sarikaya, 2015; Vielhaber et al., 2003), forming a dysfunctional neural network more extensive than the specific epileptogenic zone. Therefore, FDG-PET studies provide prognostic value in JOSÉ PAREDES PACHECO 34 the evolution of neural networks associated with epilepsy. On the other hand, the findings observed in the FDG-PET studies will also influence the post-operative status, especially in cases with limited hypometabolic extent where the probability of a good outcome is greater (Verger et al., 2018). b) Other clinical uses Regarding dementias, in MCI cases the role of FDG-PET is to evaluate the reduction in brain metabolism in relation to neural injury or synaptic dysfunction and provide information to predict whether this MCI will evolve to AD or other dementias, such as DLB or FTLD (DeCarli, 2003; Drzezga et al., 2003; Ito et al., 2015; Pagani et al., 2017). Among all types of dementia, AD is the most common, especially in subjects over 65 years of age with 50-60% of cases of dementia (Bohnen et al., 2012). Many FDG-PET studies are performed to study this disease (J. M. Hoffman et al., 2000; Mosconi et al., 2006). In Figure 28 we can see a representation of how healthy levels of metabolism decrease in cases of MCI and subsequently form different patterns of hypometabolism in AD. Figure 28: Representative FDG-PET images of a group of healthy people without cognition impairment (NCI), another group with MCI and another with AD. Adapted with permission from (Arbizu et al., 2015). On the other hand, hypometabolism patterns are characteristic of each neurodegenerative disorder (see Figure 29), leading to a fairly widespread use of FDG-PET as a tool to differentiate AD from other neurodegenerative dementias (Bohnen et al., 2012; Guillén et al., 2020; Kato et al., 2016). Introduction 41 In these cases, the importance of amyloid-PET resides in whether it is negative, since it allows to exclude AD as a possible cause of dementia (Camacho et al., 2018; Shivamurthy et al., 2015). b) Other clinical uses Other uses that are beginning to be made of amyloid-PET are in cases of PD, down syndrome, multiple sclerosis and in the evaluation of AD therapies (Camacho et al., 2018). In relation to the other imaging modalities, Amyloid-PET has shown to be better than FDGPET and MRI in detecting AD in early stages, while FDG PET is superior in predicting the progression from MCI to AD and MRI is useful to assess the severity of damage and disease progression in more advanced stages. Therefore, the combination of different modalities can lead to better prognostic values, especially when they are atypical AD situations, a strange cognitive evolution or the presence of another disease that limits a good diagnosis (Camacho et al., 2018). 1.2.3.2 Evaluation a) Quantification methods In clinical practice, when an amyloid-PET is evaluated as positive or negative, it is usually done qualitatively, that is, by visual analysis. As in the case of FDG-PET and other imaging modalities, this can sometimes lead to incorrect diagnoses or dependent on the experience of the clinician (Mountz et al., 2015; Yamane et al., 2017). For this reason, the use of quantitative methods, as a complement to the visual evaluation of amyloid-PET, has recently been promoted. One of the most widely used, both in clinical and research, is the standardized uptake value ratio (SUVr), which is calculated as the ratio between the SUV of a cortical ROI and the SUV of a reference region. Depending on the radiotracer, positive amyloid cases are usually considered when the SUVr is between 1.12 and 1.45 (Arbizu et al., 2015), although it is usually converted to the Centiloid scale (Klunk et al., 2015) that has standardized values and allows comparison between different amyloid tracers. To calculate the SUVr, PET images have to be also preprocessed. In this case, it usually consists of a spatial normalization of the images to a standard space, the manual definition of the ROI and the calculation of the SUVr (Tsubaki et al., 2020). There are several commercial software that perform this process such as PMOD (PMOD Technologies, Switzerland), HERMES Brass (Hermes Medical Solutions, Sweden), MIMneuro (MIM Software, USA) or Siemens Syngo VIA Amyloid Plaque (Siemens Medical Solutions Inc., USA) in addition to other tools under development (Tsubaki et al., 2020). However, each software has its peculiarities in the preprocessing and there is no standard in the methods for SUVr calculations (Curry et al., 2019; Knesaurek et al., 2014). b) Examples The SUVr has an important role in longitudinal studies on the evolution of amyloid deposits (Landau et al., 2015), in the detection of early stages of AD (Yamane et al., 2017) as well as in the evaluation of cognitively healthy older people where there is a low deposition of amyloid difficult to detect visually. Thanks to SUVr, this detection of less widespread betaamyloid can be improved and individuals at increased risk of having AD can be identified (Harn et al., 2017). On the other hand, SUVr has several limitations: SUVr values depend not only on the chosen reference and interest regions but also on the particular radiotracer being used (Matsuda et al., 2021). In addition, they are values that can be altered by partial volume effects. For this JOSÉ PAREDES PACHECO 42 reason, new ways to improve the preprocessing for the calculation of the SUVr are still being investigated, new quantitative measures of amyloid are being implemented, such as the Centiloid scaling method mentioned above (Klunk et al., 2015) and the AMYQ index (Pegueroles et al., 2021), or the usefulness of other analyzes are being evaluated such as VBA in amyloid-PET (Akamatsu, Ikari, et al., 2019). 1.3 MONTE CARLO SIMULATION IN BRAIN PET In this section we talk about MC simulation techniques and its different applications, particularly in nuclear medicine and brain PET imaging. In addition to the main advantages and limitations of using MC simulation compared to other methods. 1.3.1 Description MC simulation techniques are statistical methods used to solve mathematical problems, especially those that cannot be solved analytically or when experimental measurements may be impractical, (Raeside, 1976; Zaidi, 1999). In this way, the aim is to reproduce the behavior of real situations to later analyze their characteristics and study how they will evolve. These techniques have multiple applications in medical physics such as diagnostic radiology and nuclear medicine, radiotherapy physics or radiation protection (Andreo, 1991). In this case, we are going to focus on their uses in nuclear medical imaging, particularly in brain PET imaging. In PET imaging, these techniques are very useful thanks to the stochastic nature of positron emission. MC simulation allows the random generation of photons in order to study the physical processes that occur inside the scanner. This is achieved by assuming that the behavior of the system can be defined using probability density functions (PDF's). These PDF's use random numbers uniformly distributed in an interval [0,1] to generate and sample primary photons, projections, scattered photons, etc. (see Figure 35) (Zaidi, 1999). In this way, MC simulations make it possible to generate synthetic images of brain PET in a theoretical framework of reference (ground-truth) and thus evaluate different processes on controlled images (García et al., 2020). Figure 35: Physical system compared to Monte Carlo simulation. Reproduced with permission of John Wiley and Sons from (Zaidi, 1999). Introduction 43 1.3.2 Applications in brain PET MC simulations are present in many fields of neuroimaging such as evaluation of PET protocols, quantification methods, visual detectability studies, generation of datasets, among others (Andreo, 1991; Zaidi, 1999). 1.3.2.1 Evaluation of acquisition protocols and reconstruction methods MC simulation is commonly used for the evaluation of some of the different steps that are involved in the generation of the PET images, such as data acquisition and reconstruction protocols, to later establish a better configuration or improve image correction techniques. On the one hand, MC simulations are widely used to evaluate acquisition protocols and predict the performance of new PET scanner prototypes (Antonecchia et al., 2020; Guerra et al., 2018; Kochebina et al., 2019; Sheikhzadeh et al., 2017). On the other hand, MC simulations are very useful when evaluating reconstruction algorithms because the reconstructed images obtained from the different reconstruction methods under evaluation can be compared to a ground-truth image (see Figure 36). Thus, MC simulations can be used to compare different PET image reconstruction methods (Reilhac et al., 2008) or to develop new, more accurate and efficient reconstruction algorithms (Hashimoto et al., 2022; Kontaxakis, 2002) among other applications. Figure 36: Example of simulated and reconstructed images using different reconstruction methods. From left to right we have the ground truth, the images reconstructed with FBP and ML-EM, and those generated with new methods using noise and MRI information as input. Adapted with permission of IEEE from (Hashimoto et al., 2022). Regarding attenuation and scatter correction techniques in brain PET, MC simulations have an important role since they allow separating the simulated data into scattered and unscattered events and correcting these effects using an algorithm (Levin et al., 1995), as we saw in the section Scatter correction. In addition to being part of a correction method, MC simulations of PET images or different spatial distributions (see Figure 37) can also be used to evaluate different attenuation or scatter correction techniques respectively (Castiglioni et al., 1999; García et al., 2020), see the effects of scatter correction when comparing attenuation maps (Burgos et al., 2014), and more. MC simulations also have some relevance in seeing how the noise and the spatial resolution of the data affect how realistic the simulated PET images can be compared to studies obtained with real protocols in the clinical routine (Stute et al., 2011). JOSÉ PAREDES PACHECO 44 Figure 37: Attenuation maps and simulated brain PET of the same subject using different attenuation correction methods based on a CT image of the patient (A), an MRI of the patient (B) or an average CT image (C). Reproduced with permission of Sociedad Española de Física Médica from (García et al., 2020). 1.3.2.2 Evaluation and optimization of quantification methods As we have seen in previous sections, quantification methods are an important complement to visual evaluation in the study of various neurological diseases. However, the presence of quantitative analyzes in clinical routine has been diminished by the lack of standardization in quantification protocols. For this reason, MC simulations are especially useful in this task since they allow the generation of realistic images of brain PET and can be considered as ground-truth when studying different processes involved in quantification. Introduction 45 In recent years there are more and more works that use MC simulations for the study and optimization of processes involved in the quantification of PET images. Generation of images to compare correction methods for the partial-volume effect (Frouin et al., 2002); evaluation of how anatomical variability, reconstruction algorithm, special normalization algorithm, and other aspects affect sensitivity and specificity in statistical analyzes (Aguiar et al., 2008; Davatzikos et al., 2001; Ishii et al., 2001; Schoenahl et al., 2003) are just some examples from the variety of applications that simulations have in the field of PET quantification. 1.3.2.3 Visual detectability studies One type of study that is beginning to benefit from the realistic brain PET images generated by MC simulations are visual sensitivity studies in the detection of pathologies. Most detectability studies are done using a clinical ground-truth based on the response to treatments, surgeries or postmortem observations. But thanks to MC simulations, brain PET images can be generated by controlling with great precision the location and degree of development of a pathology of interest, being a better ground-truth. These works allow to study the reliability of the images acquired to show certain abnormalities in the brain, discover the detection limit of the radiotracer and study the variability in diagnosis among clinical experts. 1.3.2.4 Generation of datasets and other uses As we have seen, the MC simulation of realistic brain PET images has many applications as it allows evaluating and improving a multitude of aspects in terms of acquisition and reconstruction algorithms, quantification methods, image post-processing, etc. All the commented applications make use, to a greater or lesser extent, of realistic synthetic brain PET databases. Therefore, another proper application of MC simulation in brain PET is the ability to generate datasets for all types of works. Some datasets are available for public use by doctors or researchers (Mota et al., 2015; Papadimitroulas et al., 2013; Reilhac et al., 2005). 1.3.3 Limitations Despite all the advantages and applications that we have seen of MC simulations in brain PET, it also has some limitations. The main limitation of the MC techniques is the great computational power required to simulate a number of events that is representative of the reality of the PET scanner and that converges to the solution in a time reasonable (Buvat & Castiglioni, 2002; Zaidi, 1999). This is something that affects most experts or research groups who want to make use of these methods and who need to acquire very powerful computing equipment. On the other hand, thanks to the advances of the internet, in recent years multiple general and PET-specific MC codes have been made available to the public (Buvat & Castiglioni, 2002). However, their complexity of use and the knowledge that is needed to get the most out of it, has greatly limited its dissemination and use by researchers or health personnel who do not have the knowledge of physics, mathematics or programming required. 1.4 JUSTIFICATION PET imaging is one of the most widely used modalities in neurology to detect the early stage of a disease, diagnose its current progress, the changes associated with different treatments or therapies that have been practiced, and even predict the evolution of symptoms from the current state. This is achieved thanks to the use of different radiotracers that allow studying different functional aspects of the brain. In particular, FDG-PET is one of the most widespread modalities and provides important information about glucose metabolism in different areas of the brain. JOSÉ PAREDES PACHECO 46 The clinical assessment of PET images, in particular FDG-PET, have normally been visually evaluated, being an important support for other more traditional morphological tests such as MRI. However, as we have seen in previous sections, numerous studies have shown the importance of quantitative analysis of PET images as a complement to visual analysis and thus be able to obtain their maximum potential as tools to support the diagnosis of neurological diseases, highlighting epilepsy. FDG-PET quantification allows detecting hypometabolism in different types of epilepsy, locate epileptogenic foci even in cases of cortical epilepsy with difficult visual detection, and study metabolic asymmetry to predict how epilepsy will evolve after surgery. In fact, the European Journal of Nuclear Medicine (EANM) suggests the use of quantification methods for the study of brain PET in one of its recent clinical guidelines (Varrone et al., 2009). This has led to the emergence of different software focused on performing quantitative analysis on PET images. However, the quantitative analysis of brain PET is not still widespread in the clinical routine. It is mainly due to lack of standardization between different quantification strategies and software solutions, which can be substantially different from each other. These methods can involve different intensity normalization methods, spatial normalization transformations, anatomical templates, statistical criteria, etc. In epilepsy there are usually cases with smaller lesions that are difficult to locate visually, causing situations that are highly dependent on the quantification methods used. There are even situations, such as in many cases of TLE, where the precision of the diagnosis or the good result of the surgical operation depends on the evaluation of the hypometabolisms obtained by quantification. In addition, many of the brain PET quantification methods need databases of healthy subjects, which are not being used properly, since it is not possible to have healthy databases for each scanner model, each particular acquisition protocol or each reconstruction method. All this together leads to a huge variability between quantification outputs that should be urgently addressed. The challenge of the standardization requires having a ground truth against which each quantification strategy can be validated. One of the most common tools for the generation of ground truth values has been the use of physical phantoms, such as Hoffman phantom (E. J. Hoffman et al., 1990). However, physical phantoms have a very inflexible configuration, leading in many cases to unrealistic images of the brain, such as uniform uptake or lack of anatomical variability between subjects. An alternative tool is the use of MC simulation, which allows the generation of multiple datasets under configurable conditions. From digital phantoms, PET sinograms can be simulated and then reconstructed in order to obtain simulated PET images. Through MC simulation we can control the activity of different brain regions with greater precision and, in the case of epilepsy, generate databases with induced hypometabolism that simulate cases of cortical epilepsy. This database with a known level and location of hypometabolisms can serve as a real ground-truth for the evaluation of all types of neuroimaging protocols in epilepsy. However, the generation of useful and versatile PET databases requires the use of realistic digital phantoms, which involves additional MC simulations, greatly increasing the computing requirements. Moreover, the use of MC techniques need advanced knowledge in physics, programming, and statistics. Nowadays, there are several open MC simulation codes available for PET community, but the mentioned requirements have limited its use to researchers with solid background in MC simulation. These issues have led to the need for MC tools specially developed for non-expert researchers with limited programming skills or computational resources. This tool should include an user-friendly environment, pre-validated scanners and certain automatic processes, but keeping a degree of customization to allow a flexible generation of PET databases. Introduction 47 1.5 OBJECTIVES The main objective of the thesis is the design, development and validation of a web-based platform for the generation of simulated brain PET datasets using MC simulation techniques. 1.5.1 Specific objectives a) Validation of MC simulation models of commercial PET scanners b) Design and development of the web-based platform for MC simulation of brain PET images: SimPET. c) Generation of healthy simulated PET image datasets with different configurations for SimPET validation. d) Application of SimPET to generate a simulated PET database representative of subjects with cortical epilepsy. METHODS Methods 61 Figure 40: Image quality phantom with background ROIs. Reproduced from (National Electrical Manufacturers Association, 2007). For image quality analysis, a cross-sectional image centered on the hot and cold spheres is used. In this image, some ROIs are drawn inside the spheres and, on the other hand, other ROIs of the same size are drawn in the background of the phantom, as they appear in Figure 40. In particular, twelve 37-mm diameter ROIs are drawn in the background and ROIs of the smaller sizes (10, 13, 17, 22, and 28 mm) are drawn concentric to the 37-mm background ROIs. These same background ROIs are drawn in slices +/- 1 and +/- 2 making a total of 60 background ROIs of each size, 12 in each slice. Background variability and image contrast ratios in the spheres are the measures considered as image quality (National Electrical Manufacturers Association, 2007). The contrast ratio 𝑄𝐻,𝑗 in each hot sphere 𝑗 follows this formula: 𝑄𝐻,𝑗 =𝐶𝐻,𝑗 𝐶𝐵,𝑗 ⁄ −1 𝑎𝐻𝑎𝐵 ⁄ −1 ∗100% (17) where 𝐶𝐻,𝑗 is the average counts in the inside ROI for 𝑗, 𝐶𝐵,𝑗 is the average of the background ROI counts for 𝑗, 𝑎𝐻 is the activity concentration in all hot spheres, and 𝑎𝐵 is the activity concentration in the background. The contrast ratio 𝑄𝐶,𝑗 in each cold sphere 𝑗 is calculated by: 𝑄𝐶,𝑗 =(1−𝐶𝐶,𝑗 𝐶𝐵,𝑗)∗100% (18) where 𝐶𝐵,𝑗 is the average of the background ROI counts for 𝑗, and 𝐶𝐶,𝑗 is the average counts in the ROI for 𝑗. The percentage of background variability 𝑁𝑗 for each sphere 𝑗 follows the equation: 𝑁𝑗= 𝑆𝐷𝑗 𝐶𝐵,𝑗 ∗100% (19) JOSÉ PAREDES PACHECO 62 where 𝑆𝐷𝑗 is the standard deviation of the background ROI counts for 𝑗. In this work, the MC models of the PET scanners implemented in SimPET were tested using the previous protocols detailed in the NEMA-NU 2007 standard (National Electrical Manufacturers Association, 2007). In this way the simulated NEMA results were validated for each scanner as detailed in Chapter 3.1 of the Results. The validated MC simulation models were integrated into the web platform. 2.5 WEB-BASED PLATFORM The SimPET platform is an adaptation from Neurocloud®, a commercial online platform hosting quantification tools (https://qubiotech.com/en/neurocloud) (Qubiotech Health Intelligence SL, A Coruña, Spain). It has a multilayered architecture, with three main layers: presentation (GUI), domain logic, and data storage (see Figure 41). Each layer is built with its own technologies and can be independently upgraded, debugged, and repaired. The web portal (www.sim-pet.org) provides a simple graphical user interface (GUI). The application manages the web server, computational processes, and file storage. The platform distributes load to parallel processing “workers" that perform the simulation/reconstruction. The underlying scripts are written in various programming languages, such as Bash, MATLAB, C, and Python. Various well-validated libraries are used under-the-hood, such as SimSET, STIR, SPM12, FSL (Jenkinson et al., 2012), and the Python NiBabel (Brett et al., 2019). Figure 41: System architecture of SimPET with the three layers: Presentation, domain logic, and data storage. Methods 63 The whole system (including the web application + two simulation workers) is currently deployed on a Lenovo ThinkStation P920 with two Intel Zeon Silver 4114 processors (twenty 2.20 GHz cores and forty threads in total), 126 GB RAM and 6 TB disk space in a mirroring configuration using Ubuntu 16.04 LTS. In addition, thanks to the platform features inherited from Neurocloud, the platform is ready to be deployed in a full cloud-computing configuration, with the ability of managing autoscaling of the number of simulation cores depending on the load. 2.6 GENERATION OF HEALTHY DATABASES The validation of the platform was based on the simulation of realistic healthy patient’s FDGPET images databases for the included scanner models. To this end, FDG-PET/CT and MRI images from 25 healthy subjects were scanned with the GE Discovery ST (Group 1) and used for the phantom generation. The 25 generated activity and attenuation maps were then used as inputs for the simulation of synthetic data. 25 brain FDG-PET control subjects were simulated in SimPET using the GE Discovery ST MC model. Real noise, a whole-body injected dose of 245 MBq and a study duration of 1200 s were simulated. Scatter correction was performed during simulation of the sinograms. These sinograms were reconstructed using the OSEM reconstruction algorithm with attenuation correction, 32 iterations and 7 subsets. The images generated have a voxel size of 1x1x0.96 mm and a matrix size of 256x256x256, the same as the activity maps used as a reference. 25 brain FDG-PET control subjects have been simulated with the SimPET platform using the MC model of the GE Advance NXi scanner. Real noise, a whole-body injected dose of 370 MBq and a study duration of 1800 s were the input parameters to simulate. Scatter and attenuation correction were performed in the simulation of the sinograms. The sinograms were reconstructed with the OSEM reconstruction algorithm with 56 iterations and 14 subsets. The images generated have a voxel size of 1x1x0.96 mm and a matrix size of 256x256x256, the same as the activity maps used as a reference. Regarding the dataset with the MC model of the Siemens mCT, 23 brain FDG-PET control subjects were simulated with the SimPET platform. In this case, real noise, a whole-body injected dose of 185 MBq and a study duration of 900 s were considered as input parameters. Scatter correction was performed during simulation and the sinograms were reconstructed using the OSEM reconstruction algorithm with attenuation correction, 130 iterations and 26 subsets. As in the previous cases, the images generated have a voxel size of 1x1x0.96 mm and a matrix size of 256x256x256. The validation was carried out by performing different statistical analyzes between the simulated datasets and real subject FDG-PET images acquired on each of the scanners. 2.7 GENERATION OF AN EPILEPSY DATABASE After the validation of the platform, SimPET was used for generating a highly realistic database including different types of hypometabolisms typical of MRI-negative epilepsy. The methodology used to generate the simulated image dataset can be seen in Figure 42. The 25 GE Discovery ST activity and attenuation maps generated for the SimPET validation were used as the basis for the generation of hypometabolisms in typical regions of cortical dysplasia. In these 25 activities maps we induce different intensities hypometabolisms in randomly selected regions of the cerebral cortex, not including subcortical structures. In particular, the algorithm intersects the probabilistic brain atlas of Hammers (Hammers et al., 2003) with a gray matter probability map to extract a random ROI where to apply hypometabolism. The Hammers atlas is a label-based maximum probability map of the JOSÉ PAREDES PACHECO 64 neuroanatomy of a population and allows, among other things, the anatomical labeling of the different regions of the brain. A labeled list with the regions that can be selected can be seen in (Hammers et al., 2003). The gray matter probability map is obtained from the MRI segmentation used during the generation of each activity map. Once the ROI is selected, we consider two types of synthetic hypometabolisms that are usually present in cases of cortical dysplasia. We define type 1 hypometabolism as punctual hypometabolisms that act as epileptogenic foci within the ROI and type 2 hypometabolism as extensive hypometabolisms that usually affect a large part of the ROI. In this way, we provide the database with greater variability and diversity of hypometabolisms that reflect the real situations that would arise. In particular, we use the selected ROI to generate one of the two types of hypometabolisms considered. Type 1 hypometabolism consists of selecting a voxel within the ROI as the center of a cube with side 2 cm. This cube has a value of 1 in the center and will gradually decrease until it is 0.5 at the edges to generate a homogeneous and realistic hypometabolism in the simulated PET image. Masking the ROI with this cube and multiplying the values by a level of hypometabolism (60, 75 or 90%), we obtain a type 1 hypometabolic region. Type 2 hypometabolism consists of selecting a voxel within the ROI as the center of the hypometabolic region. In this case, the ROI will have a value of 1 in the center of the voxel and will decrease to 0 at the edges of the ROI, thus achieving a complete hypometabolic ROI. This hypometabolic ROI is multiplied by a level of hypometabolism (30 or 40%) to obtain a type 2 hypometabolic region. All these hypometabolic regions separately were applied to the activity maps (3 type 1 hypometabolic region and 1 type 2 hypometabolic region to each activity map). These activity maps with different induced hypometabolisms, together with the corresponding original attenuation maps, were simulated with SimPET (GE Discovery ST model, whole body injected dose = 245 MBq and study duration = 1200 s) to generate a dataset of 100 synthetic images with different cases of cortical epilepsy. Methods 65 Figure 42: Generation of simulated PET images with two types of induced hypometabolism from randomly selected regions. Each individual simulated image was reviewed several times by a nuclear medicine physician with previous experience in epilepsy for validating that the images and the induced metabolisms were realistic for the eyes of an expert. If it was not a realistic study, it would be generated again taking into account the comments of the nuclear medicine physician. Thus, the simulated database is refined until finally certifying that all the simulated studies form a dataset of realistic JOSÉ PAREDES PACHECO 66 cases of cortical dysplasia. In addition, the 25 brain FDG-PET control subjects simulated with the GE Discovery ST model described in chapter 2.6 were included in the final dataset. To complement the visual assessment and validate the simulated epilepsy database, a quantitative analysis was performed between the healthy simulated subjects and each of the simulated subjects with induced hypometabolism. 2.8 IMAGE PROCESSING AND QUANTIFICATION The quantitative analysis of PET images requires previous image processing, mainly in terms of spatial and intensity normalization, which can be carried out using well-known software packages such as SPM. 2.8.1 SPM SPM is a free software package created for preprocessing and statistical evaluation by hypothesis testing of different types of neuroimaging, including anatomical or functional images. This package was developed in 1994 by Karl Friston and other researchers at the Wellcome Center for Human Neuroimaging at the University College of London (Friston, 2007). Since then it has undergone different updates to the current version, SPM12. This version is the one used in this project. This software allowed us to preprocess different PET image datasets and perform VBAs. For the works present in this thesis, the preprocessing of the real or simulated PET images is similar to that described in the image analysis section of FDG-PET (see Figure 43). In particular, we used a 6-parameter rigid body transformation to co-register the acquired and simulated PET images to the space of the T1-weighted images corresponding to each subject. The PET studies were spatially normalized onto the MNI space (voxel size of 2 × 2 × 2 mm) using a 18F-FDG template and the 12-parameter affine normalization provided by SPM. This is followed by a nonlinear deformation using a mean squared difference matching function (Ashburner & Friston, 2000). A Gaussian kernel of 8 mm was used to smooth the normalized PET studies. In the intensity normalization step there is no established standard method but there are different options inside and outside of SPM that can offer good results. In our case, histogram-based intensity normalization was performed. In this method, all the smoothed PET images were voxel-wise divided by the mean of the smooth acquired images. Histogram of these ratio images were generated and each smoothed image was divided by the most prevalent value in its ratio image. Once all PET images are pre-processed, they are ready for voxel-level statistical analyzes. These VBAs allow to find the significant differences between the different groups of PET images under a certain configuration. The statistical analyzes carried out for the different works are described in the following section. Methods 67 Figure 43: Preprocessing brain PET images with SPM JOSÉ PAREDES PACHECO 68 2.9 STATISTICS In the evolution of research with PET neuroimaging, statistics have played a fundamental role, since it allows statistical inference between groups of images and to study the significant differences in the activation of different brain regions, carry out functional connectivity studies, see the functional differences in the brain before and after a treatment, study the correlation between a biological variable and the changes observed in a group, or predict the evolution of a patient, among other analyzes. Initially, the statistical analysis of PET images was based mostly on performing a voxellevel t-test between groups to study significant differences between PET images. However, later the general linear model (GLM) was also implemented in SPM as follows: 𝑦=𝑋𝛽+ 𝑒 (20) where 𝑦 are the input values of the PET images, 𝑋 the design matrix with the explanatory variables of the model, 𝛽 the regressors to be estimated and 𝑒 the errors. With the GLM defined, the regressors are estimated using ordinary least squares such that the sum of the squared errors is minimal. Once the model has been fitted and the beta regressors estimated for each voxel, statistical inference is carried out using hypothesis tests on the beta regressors. In SPM, a contrast is defined as a vector of weights that indicates the participation of each regressor in the definition of the null hypothesis. The hypothesis is then evaluated in each voxel by performing a statistical test, providing maps of t-values following a Gaussian distribution. Thus, each t-value has an associated probability value, p-value and finally a threshold (significance level, α) has to be defined to determine the voxels showing significant differences between groups. The most common α value is 0.05 (although 0.1 or 0.01 are also used). Applying this threshold we would obtain all the voxels where the null hypothesis is rejected. Adjacent voxels are grouped into clusters. These clusters form a statistical parametric map of significant differences. The minimum cluster size considered for the formation of the statistical map is defined in SPM by a parameter K. Due to the large number of voxels and t-tests that are carried out, there is the problem of multiple comparisons. If the significance level of α is set to 0.05, there is a 5% chance of providing significant differences. Considering the large number of voxels of the PET images, the proportion of false positives included in the statistical parametric map can be important. To compensate this issue, SPM implemented several correction methods, such as Bonferroni or Family Wise Error (FWE), which are applied during thresholding (Han & Glenn, 2017). In Figure 44 we can see a diagram of the entire process to obtain the statistical parametric map with SPM. In the course of this thesis, we have used the GLM implemented in SPM configured as two-sample t-test and paired t-test to compare simulated and patient brain PET images. This type of analysis allows to assess regional/systematic differences between the simulated/real groups. In addition, we have also used Bland-Altman plot to study the correlation between two groups of PET data. Methods 69 Figure 44: Generation of a statistical parametric map using the GLM with SPM. The preprocessed PET images are the input of the GLM, where 𝒚 is the intensity value in each voxel and takes a different value for each subject, 𝑿 is the design matrix that describes the classification of the images in groups, 𝜷 are the regressors to estimate and 𝒆 the errors. The 𝜷 parameters of the model are fitted by regression on each voxel taking into account the values in all subjects. Different t-tests and thresholds are applied to the estimated model to obtain statistical parametric maps. JOSÉ PAREDES PACHECO 70 2.9.1 Two-sample t-test The two-sample t-test is one of the most widely used t-tests to determine whether the mean values between two independent data groups are significantly different or not, the null hypothesis being that the mean value is the same. Applied to the VBAs, we used them to assess the significant differences at the voxel level between the mean images of two independent groups of pre-processed PET images. The two-sample t-test design matrix in SPM includes variables to indicate the group to which the input PET images belong and other optional covariates with quantitative information about the subjects such as age or sex. In our case, we directly compared the images between groups without other covariates. After the model has been configured and estimated, the contrast to be evaluated is chosen. This contrast indicates how the mean values are subtracted to obtain the significant differences for a particular comparison. Once the contrast is selected, the value from which a voxel is significantly different is determined by the significance level α. An FWE correction is applied to reduce the probability of false positives in the results. A minimum cluster size, K, for the formation of the statistical parametric map is set. In the validation of the simulated healthy databases, two different contrasts were used to assess areas with significant differences: the real group metabolism > the simulated group metabolism and the real group metabolism < the simulated group metabolism. A statistic threshold of p < 0.01 and a cluster size k = 300 was applied. In the validation of the simulated epilepsy database, the contrast simulated control group > each subject with simulated hypometabolism was used to verify the areas with induced hypometabolism. A statistic threshold of p < 0.05 and a cluster size k = 250 was applied. In any case, family wise error (FWE) correction was applied to assess for multiple comparisons. In this way, we obtain a tridimensional maps with the significantly different regions for each contrast. 2.9.2 Paired t-test The paired t-test is similar to the two-sample t-test described above, but must be used for different data structures. The paired t-test is designed to be used with pair-correlated data. For example, if we want to compare the same group of subjects with PET images generated under two different conditions or at two different times. Furthermore, while in the two-sample t-test the data have the same variance and are normally distributed, in the paired t-test the data only have to be normally distributed (Fralick et al., 2017). In the thesis we use the paired t-test to compare the real and simulated GE Discovery ST images as the activity and attenuation maps were derived from the real GE Discovery ST images. For the rest of comparisons, we used the previously commented two-sample t-test. 2.9.3 Bland-Altman comparison analysis Bland-Altman comparison analysis is a graphical method to study the differences between two measurements. This type of graphic analysis is especially useful when we study the correlation or similarity between two sets of data with respect to the same variable but acquired in a different way. It usually consists of representing the differences or ratios between two sets of data against the mean of all the data. Alternatively it can also be done with respect to one of the two data sets, if this is a reference set, instead of using the mean. In the validation of the simulated healthy databases, we adapt this analysis to study the similarity between groups of simulated PET images and real PET images for each scanner. In particular, the simulated images were compared with the real images by performing a BlandAltman-like analysis, where for each voxel, we calculated voxel-based differences (∈𝑣𝑜𝑥𝑒𝑙) as: Results 77 Table 5a: Main physical parameters implemented in the simulation model of each scanner Scanner model Scanner description Parameter Detector model Number of rings Number of detectors per ring Inner ring diameter Average depth of interaction Distance between rings Min / Max z Num aa bins Num td bins Min / Max td Min energy window Definition Detector block model type to simulate Number of rings (axial slices) Number of detectors per ring in total Inner detector ring diameter Mean distance traveled by a photon when interacting with the scintillator cristal Space between rings FOV limits in axial plane Number of azimuthal angle bins Number of transaxial distance bins Limits of transaxial distance Minimum value of the energy window GE Discovery ST Simple pet (only apply gaussian blurring) 24 420 88.62 cm 0.84 cm 0.654 cm -7.85 / 7.85 210 249 -35.12 / 35.12 375 keV GE Advance NXi Simple pet (only apply gaussian blurring) 18 672 92.7 cm 0.01 cm 0.844 cm -7.6 / 7.6 336 281 -56.2 / 56.2 375 keV Siemens Biograph mCT Simple pet (only apply gaussian blurring) 52 624 84.90 cm 0.1 cm 0.419 cm -10.9 / 10.9 312 312 -31.2 / 31.2 435 keV JOSÉ PAREDES PACHECO 78 Table 5b: Main parameters configured in the simulation/reconstruction model of each scanner. Simulation Reconstruction Parameter PSF value Add noise Apply att correction Scatt correction factor Zoom factor XY output size Reconstruction type Num iterations Num subsets Max segment Definition Value of the Point Spread Function applied to blurring the sinograms Add poisson noise to sinograms Apply attenuation correction Factor by which the scatter is corrected (0.15 is 85% corrected scatter) Zoom factor in X and Y dimensi ons X and Y sizes in reconstruc ted matrix Reconstruction algorithm Number of iterations in the reconstructi on algorithm Number of blocks into which the image reconstructi on is divided Maximum ring difference on the reconstruction GE Discovery ST 1.125 - Yes 0.15 1.23 128 OSEM 3D 32 7 23 GE Advance NXi 1.125 0.01 Yes 0.15 2 128 OSEM 3D 56 14 11 Siemens Biograph mCT 1 - Yes 0.15 1 200 OSEM 3D 130 26 49 Results 79 3.2 SIMPET - AN OPEN ONLINE PLATFORM FOR THE MONTE CARLO SIMULATION OF REALISTIC BRAIN PET DATA 3.2.1 SimPET platform After the validation of the MC simulation models in the three commercial scanners, the subsequent result of this thesis was the creation of the SimPET platform. In this chapter we present SimPET (www.sim-pet.org), a free, easy-to-use, cloud-based platform for the generation of synthetic PET images with a special focus on brain imaging. The source code is free and can be downloaded from GitHub (https://github.com/txusser/brainviset_simset). The SimPET collaboration is an initiative of the Health Research Institute of Santiago (IDIS) with partners such as the General Foundation of the University of Malaga (FGUMA), University of Barcelona (UB) and Qubiotech Health Intelligence SL. The platform allows the automatic generation of realistic digital brain phantoms derived from patient PET/CT and MRI images by using SimSET and STIR within the Brain-VISET (voxel-based iterative simulation for emission tomography) method previously published by our group (Marti-Fuster et al., 2014), and the simulation and reconstruction of these or other user-defined phantoms using the validated scanner models included in the platform. The interested reader can find past use cases in our published work (Silva-Rodríguez et al., 2014, 2015, 2016). 3.2.2 GUI and functionalities Figure 46 shows the GUI, that users can access freely through www.sim-pet.org, after filling a registration form with their email and some additional information. The main menu is structured in three modules: Phantom generation (map generation), simulation, and reconstruction. After each process, intermediate files can be downloaded/uploaded, allowing the user to introduce extra steps in the process. An interactive online image viewer is also included for evaluating the results without the need of downloading any file. The different boxes in Figure 46 provide an overview of the different interaction panels in the platform. The user profile, an online manual and online support can also be found in the main page. The platform accepts inputs in the Digital Imaging and Communications in Medicine (DICOM) and NifTI-1 formats. When uploading DICOM, the images are anonymized and converted to NifTI-1, which is a more convenient format for mathematical manipulation. The platform outputs the generated digital brain phantoms (activity and attenuation maps), sinograms, and reconstructed PET images in NifTI-1. JOSÉ PAREDES PACHECO 80 Figure 46: Web-based Graphical User Interface of SimPET (www.sim-pet.org). Surrounding boxes show the pop-up menus for patient-derived map generation (green), Monte Carlo simulation (dark blue), Tomographic reconstruction (yellow) and we highlight the different support options (light blue). A typical workflow is illustrated in Figure 47, including (i) loading PET, CT, and MR images of the same subject, as input parameters, and configuring a scanner model, a radiotracer, injected dose (MBq), and scan duration (s) for generating activity and attenuation maps using Brain-VISET; (ii) generating simulated sinograms from previously generated maps or from uploaded attenuation and activity maps using SimSET. For these, the user must set the desired injected dose (MBq), the scan duration (s), and the noise level, which can be allowed to easily generate noise-free simulations independently of the selected parameters; and finally (iii) reconstructing the above sinograms using STIR. Once the different images are generated, these can be viewed online or downloaded. For more details or an usage example, see the online manual at www.sim-pet.org (Figure 46, light blue box). Results 81 Figure 47: Typical workflow of SimPET. (i) First PET/CT and MRI images are considered as inputs for the activity and attenuation map generation. (ii) Next, the attenuation and activity maps are simulated for obtaining PET sinograms. (iii) Finally, the reconstructed PET images are generated by reconstructing these sinograms. The simulation and reconstruction steps can be repeated after changing the input parameters until the final image is satisfactory. 3.2.3 Internal processing The MC simulation is performed using SimSET (v.2.9.2), which includes the simulation of all the physical processes for the energies of interest in nuclear medicine (below 1 MeV) (R. Harrison et al., 2008; R. E. Schmitz et al., 2007). The generated data are then reconstructed by using STIR (v.3.1) (https://github.com/UCL/STIR). Currently, three commercial PET scanners are supported, namely the GE Discovery ST (GE Healthcare, Chicago, United States), The GE Advance NXi (GE Healthcare, Chicago, United States) and the Siemens Biograph mCT (Siemens Healthineers, Erlanger, Germany). Additional scanners will be progressively added in the future. Part of the parameters of the models of the three implemented scanners are based on previously published works (Barret et al., 2005; Guerin & El Fakhri, 2008; Poon et al., 2015). The scanner models are validated as shown in the results of section 3.1. The images are JOSÉ PAREDES PACHECO 82 reconstructed using the ordered subsets expectation maximization (OSEM) algorithm as implemented in STIR, setting the reconstruction parameters to match those set in the scanner as closely as possible. Realistic activity and attenuation maps including nonuniform activities and physiological variability can be fully automatically extracted from patient images by using the Brain-VISET iterative method. This method has been previously presented in detail (Marti-Fuster et al., 2014) and can be seen schematically in the orange box of Figure 47. In brief, PET/CT images are coregistered to the T1-weighted MRIs using FLIRT (https://fsl.fmrib.ox.ac.uk/fsl). Bone tissue images are extracted from the CTs using a 600 Hounsfield unit threshold. The T1 MRI is segmented to gray matter, white matter, and cerebrospinal fluid using SPM12. An initial activity map is generated by filling the segmented tissues with uniform activities. An initial attenuation map is created joining bone tissue image (attenuation value = 3) with outskin image (attenuation value = 4). The bone tissue image is generated from thresholded CT and the outskin image came from joining all segments generated with SPM12. The initial maps are then simulated using the MC model for the selected scanner. Postsimulation, the reconstructed image is compared with the original PET image studying the correlation coefficient. In case the correlation coefficient is < 0.99, the original PET image is divided by the reconstructed image to generate a ratio image and the activity map is multiplied by this ratio image to be updated. This process is repeated iteratively until the correlation coefficient is ≥ 0.99. 3.3 GENERATION OF HEALTHY DATABASES For SimPET validation, realistic simulated databases of healthy subjects were generated from digital phantoms for each of the scanners implemented in the SimPET platform: the GE Discovery ST, the GE Advance NXi and the Siemens mCT. The CPU time for the generation of digital phantoms was about 3–4 h for GE Discovery ST, reaching convergence after four or five BrainVISET iterations. Figure 48 shows a subset of Group 1 (row 1) and the corresponding generated digital phantoms (row 2). Figure 48: Sample of ten healthy control PET images acquired on GE Discovery ST (Group 1), the corresponding generated digital phantoms, and the corresponding simulated PET images obtained in the different commercial scanners using SimPET. Results 83 The computation time for the simulation of the generated phantoms was about 50 min per subject, while the reconstruction time was 20 min for the GE Discovery ST and GE Advance NXi, and 40–60 min for the Siemens Biograph mCT scanner. This difference in reconstruction times was caused by the difference in times for the calculation of attenuation sinograms between the different scanners. Figure 48 (rows 3–5) shows the corresponding simulations of this subset of Group 1 in the three scanners. On the other hand, the results of the Bland-Altman analysis are shown in Figure 49. Figure 49: Bland-Altman comparison between the real and simulated databases, for the GE Discovery ST (left), the GE Advance (center) and the Siemens mCT (right). In the left blue histogram, we can observe the differences between the real and the simulated images for the GE Discovery ST, showing that 83.02% of voxels showed differences of < 5%, while 98.09% were below 10%. For the GE Advance NXi (center), the blue histogram represents the differences between the real and the simulated images, showing that 71.98% of the voxels had differences lower than 5%, while 95.09% were below 10%. These results were similar when comparing between real databases for the GE Discovery ST (Group 1) and the GE Advance NXi (Group 2). The red histogram shows that 68.42% voxels had differences lower than 5% while 94.55% are below 10%, pointing to the fact that most of the observed differences between the real and the simulated images for the GE Advance NXi, they were due to differences between the real databases. Finally, for the Siemens Biograph mCT (right), the blue histogram shows that 62.22% of the voxels showed differences of less than 5%, while 91.35% were below 10%. Instead, when comparing between real databases for GE Discovery ST (Group 1) and Siemens Biograph mCT (Group 3), the red histogram shows differences lower than 5% for 79.03% of the voxels, and below 10% for 97.60%, showing that the model of the Siemens Biograph mCT provides the worst correspondence along the used scanner models. Figure 50 shows the voxel-wise statistical comparisons (mean and statistical differences) between acquired and simulated images for the different scanners. As it can be observed in Figure 50 (rows 1–2), good visual agreement between the mean simulated and acquired images was achieved. Voxel-wise analysis (Figure 50, row 3) showed some regional differences, including regions of higher activity in the right temporal lobe of the simulated group and a bilateral region showing lower activity for the simulated group in the upper frontal lobe. As for the GE Discovery ST, a good visual agreement between simulated and real images is observed for GE Advance NXi (Figure 50, rows 4–5). The voxel-wise analysis showed some areas of higher metabolism for the simulated group on the frontal lobe, while reduced activity was observed in the temporal lobe and the internal structures (Figure 50, row 6). For the Siemens Biograph mCT we observed the greater visual differences between the simulations and the JOSÉ PAREDES PACHECO 84 acquisitions (Figure 50, rows 7–8). Artifacts can be observed in the simulated images, such as slight asymmetries on the occipital and temporal lobes that are not observed for the other scanners. The voxel-wise analysis showed higher simulated activity in the cerebellum, the internal structures and some small cortical areas in the right frontal lobe and the left occipital lobe. In contrast, we observed reduced simulated metabolism in several small clusters distributed along the cortex. The three simulated databases are freely available via SimPET web (www.sim-pet.org) in the Datasets section. Figure 50: Voxel-based analysis results for the comparison between the simulated images and real FDGPET images for the three simulated scanners. A threshold of P < 0.01 corrected and k = 300 is applied. 3.4 GENERATION OF AN EPILEPSY DATABASE Initially, three levels of type 1 hypometabolism were considered for each activity map. Thus, 25 maps with an induced hypometabolism of 60%, 25 maps with a hypometabolism of 75% and 25 maps with a hypometabolism of 90% were created and then a total of 75 simulated images were obtained, which were visually validated by the experienced nuclear physician. This visual validation resulted in a significant proportion of non-realistic hypometabolisms that Results 85 were discarded because they showed too subtle epileptogenic foci, undetected or areas of low metabolic activity that could be confused with the epileptogenic foci. This process of review and simulation of new studies was repeated several times until we obtained a first database visually validated by our nuclear medicine physician. In this way, we ended up with 25 PET studies with an induced hypometabolism of 60%, 31 studies with a hypometabolism of 75% and 19 studies with a hypometabolism of 90%. All hypometabolisms in small random regions of the cerebral cortex (type 1 hypometabolisms). Following the recommendations of our epilepsy expert to increase the variability of the dataset, we decided to include 25 additional cases with 2 levels of type 2 hypometabolism. Thus, after some review by our expert, 15 PET studies with an induced hypometabolism of 40% and 10 studies with a hypometabolism of 30% were simulated. These hypometabolisms affect whole regions of the cerebral cortex (hypometabolism type 2). If we add the 25 simulated PET studies without induced hypometabolism to the dataset simulated with hypometabolism, we form the definitive database with a total of 125 simulated images (100 simulated PET images with induced hypometabolism + 25 simulated PET images of healthy controls). The characteristics of each of the 100 simulated PET studies with induced hypometabolism can be seen in Table 6a and Table 6b. JOSÉ PAREDES PACHECO 86 Table 6a: Characteristics of simulated PET studies with induced hypometabolism in the cerebral cortex. Subject Type of hypometabolism % of hypometabolism Hypometabolism región (atlas of Hammers) 1 Type 1 60 7-Anterior temporal lobe lateral part 2 Type 1 60 30-Posterior temporal lobe 3 Type 1 75 20-Insula 4 Type 1 60 49-Precentral gyrus 5 Type 1 75 55-Inferior frontal gyrus Brocca’s area 6 Type 1 90 3-Amygdala 7 Type 1 60 56-Inferior frontal gyrus Brocca’s area 8 Type 1 60 14-Middle and inferior temporal gyrus 9 Type 1 90 32-Inferiolateral remainder of parietal 10 Type 1 75 64-Lingual gyrus 11 Type 1 75 13-Middle and inferior temporal gyrus 12 Type 1 75 25-Gyrus cinguli posterior part 13 Type 1 75 49-Precentral gyrus 14 Type 1 75 65-Cuneus 15 Type 1 90 7-Anterior temporal love lateral part 16 Type 1 75 62-Precuneus 17 Type 1 75 26-Gyrus cinguli posterior part 18 Type 1 90 64-Lingual gyrus 19 Type 1 75 4-Amygdala 20 Type 1 75 1-Hippocampus 21 Type 1 90 16-Fusiform gyrus 22 Type 1 60 53-Anterior orbital gyrus 23 Type 1 75 56-Inferior frontal gyrus Brocca’s area 24 Type 1 75 55-Inferior frontal gyrus Brocca’s area 25 Type 1 60 60-Postcentral gyrus 26 Type 1 90 63-Lingual gyrus 27 Type 1 75 12-Temporal superior posterior part 28 Type 1 60 58-Superior frontal gyrus 29 Type 1 75 32-Inferiolateral remainder of parietal 30 Type 1 90 58-Superior frontal gyrus 31 Type 1 60 5-Anterior temporal lobe medial part 32 Type 1 75 30-Posterior temporal lobe 33 Type 1 90 13-Middle and inferior temporal gyrus 34 Type 1 60 64-Lingual gyrus 35 Type 1 60 13-Middle and inferior temporal gyrus 36 Type 1 90 54-Anterior orbital gyrus 37 Type 1 60 7-Anterior temporal love lateral part 38 Type 1 75 27-Middle frontal gyrus 39 Type 1 75 6-Anterior temporal lobe medial part 40 Type 1 60 59-Postcentral gyrus 41 Type 1 75 50-Precentral gyrus 42 Type 1 90 1-Hippocampus 43 Type 1 60 50-Precentral gyrus 44 Type 1 75 11-Temporal superior posterior part 45 Type 1 90 56-Inferior frontal gyrus Brocca’s area 46 Type 1 60 28-Middle frontal gyrus 47 Type 1 90 2-Hippocampus 48 Type 1 90 14-Middle and inferior temporal gyrus 49 Type 1 60 20-Insula 50 Type 1 75 56-Inferior frontal gyrus Brocca’s area Discussion 93 resolution. This is a pretty good result considering that the Siemens Biograph mCT is a more modern scanner than the other two and with new technologies for data acquisition, such as Time of Flight (Jakoby et al., 2011), which increase the complexity of the simulation. Considering the simulation models of the three commercial PET scanners evaluated for SimSET and fixing the physical parameters that describe the geometry of each scanner, different configurations of number of iterations in the reconstruction, PSF values, etc. were tested looking to reduce the differences obtained between the real and the simulated scanners. These tests were necessary since some of these parameters were not known a priori, and the manufacturers of the PET scanners do not usually provide them with precision. An individual improvement in the results of spatial resolution, sensitivity or image quality of the simulated scanner could be achieved using values other than those in Table 5a and Table 5b. However, improving the results in one of the tests could worsen the results in another. For example, for the GE Discovery ST simulation model, the number of iterations selected for the reconstruction minimized the errors in the NEMA tests with respect to the real scanner in general, since if we lower this parameter, the difference in the contrasts of the spheres of the image quality test decreases a little but increases the difference in background variability. On the other hand, if we increase the number of iterations, the difference in the image quality test increases in general. Also, applying lower or higher values of PSF could increase the difference in one component of the spatial resolution and decrease in the others or directly increase the difference in the whole spatial resolution. Taking all of the above into account, the configurations shown in Table 5a and Table 5b provide a better balance and lower global differences between the real and simulated scanners in the different NEMA tests, considering the MC simulation models validated under the conditions described in chapter 3.1. 4.2 SIMPET - AN OPEN ONLINE PLATFORM FOR THE MONTE CARLO SIMULATION OF REALISTIC BRAIN PET DATA The evaluation and standardization of quantification methods in brain PET require the analysis of large pools of PET images, ideally against a well-known ground truth. While physical phantoms (Akamatsu et al., 2019; Harries et al., 2020; E. J. Hoffman et al., 1990; Iida et al., 2013; Wong et al., 1984) provide a way forward, they have quite a few limitations. They are complicated devices to fill properly, so acquiring many phantom images can be tedious and time consuming. They are often limited to simple geometric shapes. The generated studies do not show variability in the uptake of WM and GM since they are uniformly filled. Even the materials used in the phantom making have a limited ability to reproduce tissue characteristics such as density or attenuation (Filippou & Tsoumpas, 2018). All this leads to acquiring unrealistic images. Therefore, simulations can be of value in this case. In recent works, easyto-use platforms for analytical simulation were presented (Pfaehler et al., 2018). The analytic PET simulators have low computational time costs and are especially useful in evaluation works of image processing methods and, in general, in works that require generating a large number of images. However, analytical simulations cannot be used, for example, to evaluate detector designs due to their low accuracy. This has led to most PET simulators being based on MC methods. MC methods allow to simulate the processes of acquisition and generation of projection data in a much more precise way, as well as to model different scanner designs and detector materials. However, great technical expertise and high computing power are required. To overcome these limitations without losing the precision provided by MC methods, we expanded the range of simulation software available by offering a free, intuitive, and efficient tool for MC simulation of realistic PET images, with a focus on brain imaging. SimPET is the JOSÉ PAREDES PACHECO 94 first web platform to have user-friendly tools for generating realistic activity and attenuation maps from patient data, simulating sinograms by combining previous generated digital phantoms with well-validated scanner models as well as reconstructing realistic synthetic brain PET images from the previously simulated sinograms. In addition, the modular design of the GUI allows the user to repeat any of the steps using different configurations, saving time and making it easy to customize simulated databases without restarting the entire process. Downloading different files of interest generated during the simulation, as well as previewing some of them through an online viewer are some of the additional features that the GUI has. SimPET is a project in continuous development and is open, since all the simulation and reconstruction code is fully available on Github (https://github.com/txusser/brainviset_simset), allowing other users to contribute new features. Currently, new more detailed detector models, such as "cylindrical" model are being tested and implemented, as well as more sophisticated features not included so far, such as time-of-flight, resolution recovery and a list-mode for reconstruction (Efthimiou et al., 2019; Karakatsanis et al., 2016; Surti, 2015). These features are currently under development and will be available on a future release (Silva-Rodríguez et al., 2019). In addition to these features, the modular architecture of the platform favors collaboration with different centers for the incorporation of new models of commercial PET scanners. These scanners can have very different geometries, from preclinical equipment to total-body PET scanners, passing through the most current PET scanners. Many of these scanners already have a validated MC simulation model (Ahmed et al., 2020; Cañadas et al., 2011; Lu et al., 2016; Monnier et al., 2015; Schmidtlein, Kirov, et al., 2006), which would facilitate their future implementation on the platform. Our future goal is to build an open and diverse scientific community of users and developers. The platform currently has more than 80 registered users since it was published online in 2020. 4.3 GENERATION OF HEALTHY DATABASES To demonstrate the capabilities of the platform, we generated synthetic databases of healthy patients for the three scanner models included in the library. For this, we generated realistic activity and attenuation maps derived from 25 healthy patients acquired on the GE Discovery ST, which were then simulated using different scanner models. Despite the images generated by the platform were visually close to the real images for all scanners (see Figure 48), statistical analysis revealed significant differences, especially for the GE Advance and the Siemens mCT. In this regard, quantitative analysis was performed by comparing simulated images with the real ones acquired on each scanner by Bland-Altman analysis and SPM statistical analysis (see Figure 49 and 50). In both cases, the GE Discovery ST showed the smallest differences between the acquired and simulated images (83.02% of voxels have differences of < 5%). This is not surprising, as this was the most favorable comparison, since the simulated images were compared with the FDG-PET images from which the digital phantoms were derived. Results from the other scanners (GE Advance NXi and Siemens Biograph mCT) showed bigger differences. Unlike the GE Discovery ST, where the differences observed are due solely to the quality of the simulation, the differences obtained for GE Advance NXi and Siemens Biograph mCT were due to both the simulation and the physiological differences between the control groups (see Figure 49) because different subjects were compared and the size of each group was relatively small. In addition, this validation had the handicap that the simulation images of the three scanners were generated from the same activity and attenuation maps. More work will be needed to decrease the variability between the databases and new features are being developed to improve MC simulation models in order to minimize the differences (SilvaRodríguez et al., 2019). Despite the differences, the tests supported the idea that simulated PET Discussion 95 studies provided good realistic results according to the simulation complexity of each implemented scanner. In this way, the simulation models currently implemented in the platform were considered validated. These databases can serve as a control group in the study of many pathological cases as well as a test bench for the evaluation of PET neuroimaging protocols, eliminating the need to acquire or have a database of real controls. In addition, they are easily expandable through the platform thanks to the implemented PET scanner models. Our simulated healthy databases are an important addition to the low availability of simulated databases of realistic brain PET studies (Batan et al., 2004; Mota et al., 2015; Reilhac et al., 2005) and they have some advantages over other datasets. Our databases were validated by different statistical analyses and respect the physical characteristics of the implemented commercial PET scanners, which makes them especially realistic and useful in works involving these or other commercial scanners. Furthermore, our datasets have a realistic variability of activity values since the digital phantoms used in the simulation were generated by the Brain-VISET iterative method (MartiFuster et al., 2014). All this makes our healthy databases an important contribution to the development of PET scanners, clinical diagnosis and research. 4.4 GENERATION OF AN EPILEPSY DATABASE As a central element of this thesis focused on the validation of new PET protocols for epilepsy, we proceeded to simulate a cortical epilepsy database using the previously generated activity maps for the GE Discovery ST scanner model. We induced activity maps with different levels and types of known hypometabolism in ROIs of the cerebral cortex to then generate simulated PET studies of cortical epilepsy. It is a unique database, both for the variety of cases of cortical epilepsy present with hypometabolism of various sizes, levels of intensity and affected regions, as well as for the visual review of each case by our experienced nuclear medicine physician and the subsequent quantitative validation. Currently there are few works that simulate PET studies with realistic hypometabolism in neurological diseases and, in particular, in epilepsy. (Baete et al., 2004) simulates hypometabolisms in the cerebral cortex from a digital phantom to evaluate the performance of a reconstruction algorithm for the detection of these hypometabolic regions in epilepsy. However, they use a digital phantom without variability in the uptake of WM and GM and perform the simulation without using a validated model or software, leading to unrealistic images. Contrary to simulated studies from our epilepsy database. This highlights the quality of the simulated database as well as the ability of the SimPET platform to simulate realistic pathological PET studies. The first application we have proposed for our database is to evaluate the visual detectability of FDG-PET in patients with epilepsy. In this type of work, the visual sensitivity of different experts in epilepsy to detect epileptogenic foci is evaluated. Most detectability studies in epilepsy are done without simulation and, therefore, without real ground-truth. Instead, the clinical history of the patients, an operative report after surgery or other medical tests are used as the reference standard (Kikuchi et al., 2021; Paldino et al., 2017; Struck & Westover, 2015). However, this entails a series of limitations such as the introduction of some patient selection bias, the possible non-existence of hypometabolism despite the success of surgery or directly the lack of knowledge of the hypometabolism that really exists since all these references are subject to errors or interpretations. With SimPET we do not have these limitations since it allows the simulation of epilepsy cases in a controlled environment, thus generating the described database. Our simulated epilepsy database constitutes a real and known ground-truth for the evaluation of epilepsy neuroimaging protocols. It is currently undergoing an international detectability study with nuclear medicine physicians from different JOSÉ PAREDES PACHECO 96 countries (Spain, France and Italy) with the aim of assessing the ability of visual analysis to locate epileptogenic foci in the cerebral cortex compared to quantitative analysis. To this end, a platform has been created from which nuclear medicine physicians can access the epilepsy database, download the quantitative analysis if necessary, and send a report with the region of the brain in which they have detected hypometabolism and the degree of confidence in the selection. Once we have all the reports from the experts, the sensitivity and specificity will be analyzed. Another application could be as a visual reference for the creation of epilepsy guidelines, taking advantage of the wide variety of cortical hypometabolisms in the database. Also as a tool to train nuclear physicians with simulated PET images with induced hypometabolism giving them the ability to accurately locate epileptogenic foci. For example, a study with several phases of training could be proposed, starting with extensive hypometabolisms as easy level and, little by little, increasing the difficulty with lighter or smaller hypometabolisms. If the database is expanded to many more cases, it could serve not only as training for nuclear physicians but also as a source of information for training machine learning algorithms. These types of algorithms need a large amount of pre-classified training data to work accurately. This is a major drawback in epilepsy works dependent on real PET studies, since there is usually a limited number of images available and it is difficult and time consuming to locate and delimit epileptogenic foci. These problems can be overcome thanks to simulated, extensible and validated databases such as the one presented in this thesis. Our database will allow training machine learning algorithms for automatic detection of hypometabolic cortical regions, classifiers of different types of hypometabolism associated with epilepsy or predictors of surgical outcome, among other uses. CONCLUSIONS 5 CONCLUSIONS We have successfully developed a new computer simulation web platform designed for brain PET studies and validated in FDG-PET images, specifically in epilepsy patients. For this, we have achieved different milestones summarized in the following conclusions: a) As an initial step, MC simulation models of three commercial PET scanners were validated and configured to be implemented into a simulation web platform. The validation of the GE Discovery ST scanner was carried out using our NEMA measurements, while the validation of GE Advance and mCT Siemens were derived from works previously developed in our group. All of them resulted in small differences between the simulated model and the real scanners. b) The simulation web platform to integrate all MC simulation models was developed as a free, user-friendly and efficient tool that allows the automatic generation of realistic digital brain phantoms and the simulation of brain PET studies. It was validated through the simulation of three databases of healthy controls, one for each implemented commercial scanner, showing an excellent agreement between simulated and real brain PET studies. c) Finally, the simulation web platform was efficiently used to generate simulated brain PET images of epilepsy patients in order to evaluate different protocols in epilepsy. A new methodology was developed to induce known and controlled hypometabolic cortical areas in digital brain phantoms and generate a well-validated database of simulated FDG-PET images of epilepsy patients. BIBLIOGRAPHY Filho, J. L. G. C., Lima, R. de S. L., Neto, L. de S. M., Bittencourt, L. K., Domingues, R. C., & Fonseca, L. M. B. da. (2011). PET/CT and vascular disease: Current concepts. European Journal of Radiology, 80(1), 60–67. https://doi.org/10.1016/j.ejrad.2010.12.102 Filippou, V., & Tsoumpas, C. (2018). Recent advances on the development of phantoms using 3D printing for imaging with CT, MRI, PET, SPECT, and ultrasound. Medical Physics, 45(9), e740–e760. https://doi.org/10.1002/mp.13058 Finocchiaro, P., Campisi, A., Corso, D., Cosentino, L., Fallica, G., Lombardo, S., Mazzillo, M., Musumeci, F., Piazza, A., Privitera, G., Privitera, S., Rimini, E., Sanfilippo, D., Sciacca, E., Scordino, A., & Tudisco, S. (2006). Test of scintillator readout with single photon avalanche photodiodes. Nuclear Science, IEEE Transactions On, 52, 3040–3046. https://doi.org/10.1109/TNS.2005.862912 Fralick, D., Zheng, J. Z., Wang, B., Tu, X. M., & Feng, C. (2017). The Differences and Similarities Between TwoSample T-Test and Paired T-Test. Shanghai Archives of Psychiatry, 29(3), 184–188. https://doi.org/10.11919/j.issn.1002-0829.217070 Frankó, E., Joly, O., & Initiative, for the A. D. N. (2013). Evaluating Alzheimer’s Disease Progression Using Rate of Regional Hippocampal Atrophy. PLOS ONE, 8(8), e71354. https://doi.org/10.1371/journal.pone.0071354 Friston, K. J. (2007). Statistical parametric mapping: The analysis of functional brain images. Elsevier / Academic Press. Frouin, V., Comtat, C., Reilhac, A., & Grégoire, M.-C. (2002). Correction of Partial-Volume Effect for PET Striatal Imaging: Fast Implementation and Study of Robustness. Journal of Nuclear Medicine, 43(12), 1715– 1726. García, J. P. C., Sala-Llonch, R., Fernández, R. T., Puig, D. R., Segura, J. P., & Baizán, A. N. (2020). Corrección de atenuación en equipos PET-RM. Comparación de métodos mediante simulación Monte Carlo. Revista de Física Médica, 21(2), 43–52. https://doi.org/10.37004/sefm/2020.21.2.004 Geramifar, P., Ay, M. R., Shamsaie Zafarghandi, M., Loudos, G., & Rahmim, A. (2009). Performance comparison of four commercial GE discovery PET/CT scanners: A monte carlo study using GATE. Iranian Journal of Nuclear Medicine, 17(2), 26–33. Ghabrial, A., Franklin, D. R., & Zaidi, H. (2021). A Monte Carlo simulation study of scatter fraction and the impact of patient BMI on scatter in long axial field-of-view PET scanners. Zeitschrift Für Medizinische Physik, 31(3), 305–315. https://doi.org/10.1016/j.zemedi.2021.01.006 Guerin, B., & El Fakhri, G. (2008). Realistic PET Monte Carlo Simulation With Pixelated Block Detectors, Light Sharing, Random Coincidences and Dead-Time Modeling. IEEE Transactions on Nuclear Science, 55(3), 942–952. https://doi.org/10.1109/TNS.2008.924064 Guerra, A. D., Ahmad, S., Avram, M., Belcari, N., Berneking, A., Biagi, L., Bisogni, M. G., Brandl, F., Cabello, J., Camarlinghi, N., Cerello, P., Choi, C.-H., Coli, S., Colpo, S., Fleury, J., Gagliardi, V., Giraudo, G., Heekeren, K., Kawohl, W., … Consortium, T. (2018). TRIMAGE: A dedicated trimodality (PET/MR/EEG) imaging tool for schizophrenia. European Psychiatry, 50, 7–20. https://doi.org/10.1016/j.eurpsy.2017.11.007 Guillén, E. F., Rosales, J. J., Lisei, D., Grisanti, F., Riverol, M., & Arbizu, J. (2020). Current role of 18F-FDG-PET in the differential diagnosis of the main forms of dementia. Clinical and Translational Imaging, 8(3), 127– 140. https://doi.org/10.1007/s40336-020-00366-0 Halbleib, J. A., Kensek, R. P., Valdez, G. D., Mehlhorn, T. A., Seltzer, S. M., & Berger, M. J. (1993). ITS Version 30: The Integrated TIGER Series of coupled electron/photon Monte Carlo transport codes (p. 7). Hammers, A., Allom, R., Koepp, M. J., Free, S. L., Myers, R., Lemieux, L., Mitchell, T. N., Brooks, D. J., & Duncan, J. S. (2003). Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe. Human Brain Mapping, 19(4), 224–247. https://doi.org/10.1002/hbm.10123 Han, H., & Glenn, A. L. (2017). Evaluating Methods of Correcting for Multiple Comparisons Implemented in SPM12 in Social Neuroscience fMRI Studies: An Example from Moral Psychology (p. 129734). https://doi.org/10.1101/129734 Harn, N. R., Hunt, S. L., Hill, J., Vidoni, E., Perry, M., & Burns, J. M. (2017). Augmenting Amyloid PET Interpretations with Quantitative Information Improves Consistency of Early Amyloid Detection. Clinical Nuclear Medicine, 42(8), 577–581. https://doi.org/10.1097/RLU.0000000000001693 Harries, J., Jochimsen, T. H., Scholz, T., Schlender, T., Barthel, H., Sabri, O., & Sattler, B. (2020). A realistic phantom of the human head for PET-MRI. EJNMMI Physics, 7(1), 52. https://doi.org/10.1186/s40658020-00320-z Harrison, R., Gillispie, S., Schmitz, R., & Lewellen, T. (2008). Modeling block detectors in SimSET. Journal of Nuclear Medicine: Official Publication, Society of Nuclear Medicine, 49 Suppl 1, 410P. Harrison, R. L., Gillispie, S. B., & Lewellen, T. K. (2006). Design and Implementation of a Block Detector Simulation in SimSET. IEEE Nuclear Science Symposium Conference Record. Nuclear Science Symposium, 5, 3151–3153. https://doi.org/10.1109/NSSMIC.2006.356543 Hashimoto, F., Ote, K., & Onishi, Y. (2022). PET Image Reconstruction Incorporating Deep Image Prior and a Forward Projection Model. IEEE Transactions on Radiation and Plasma Medical Sciences, 1–1. https://doi.org/10.1109/TRPMS.2022.3161569 Heinen, L., Teune, L., & Leenders, N. (2013). Het cerebrale glucose-metabolisme bij neurodegeneratieve hersenaandoeningen. Tijdschrift Voor Neuropsychiatrie En Gedragsneurologie, 1, 51–59. https://doi.org/10.1007/s40533-013-0011-9 Hirayama, H., Namito, Y., Bielajew, A., Wilderman, S., & Nelson, W. (2006). The EGS5 code system. https://doi.org/10.2172/877459 Hoffman, E. J., Cutler, P. D., Digby, W. M., & Mazziotta, J. C. (1990). 3-D phantom to simulate cerebral blood flow and metabolic images for PET. IEEE Transactions on Nuclear Science, 37(2), 616–620. https://doi.org/10.1109/23.106686 Hoffman, J. M., Welsh-Bohmer, K. A., Hanson, M., Crain, B., Hulette, C., Earl, N., & Coleman, R. E. (2000). FDG PET imaging in patients with pathologically verified dementia. Journal of Nuclear Medicine: Official Publication, Society of Nuclear Medicine, 41(11), 1920–1928. Hudson, H. M., & Larkin, R. S. (1994). Accelerated image reconstruction using ordered subsets of projection data. IEEE Transactions on Medical Imaging, 13(4), 601–609. https://doi.org/10.1109/42.363108 Humm, J. L., Rosenfeld, A., & Del Guerra, A. (2003). From PET detectors to PET scanners. European Journal of Nuclear Medicine and Molecular Imaging, 30(11), 1574–1597. https://doi.org/10.1007/s00259-0031266-2 Iida, H., Hori, Y., Ishida, K., Imabayashi, E., Matsuda, H., Takahashi, M., Maruno, H., Yamamoto, A., Koshino, K., Enmi, J., Iguchi, S., Moriguchi, T., Kawashima, H., & Zeniya, T. (2013). Three-dimensional brain phantom containing bone and grey matter structures with a realistic head contour. Annals of Nuclear Medicine, 27(1), 25–36. https://doi.org/10.1007/s12149-012-0655-7 Ishii, K., Willoch, F., Minoshima, S., Drzezga, E., Ficaro, E. P., Cross, D. J., Kuhl, D. E., & Schwaiger, M. (2001). Statistical brain mapping of 18F-FDG PET in Alzheimer’s disease: Validation of anatomic standardization for atrophied brains. Ito, K., Fukuyama, H., Senda, M., Ishii, K., Maeda, K., Yamamoto, Y., Ouchi, Y., Ishii, K., Okumura, A., Fujiwara, K., Kato, T., Arahata, Y., Washimi, Y., Mitsuyama, Y., Meguro, K., Ikeda, M., & Group, S.-J. S. (2015). Prediction of Outcomes in Mild Cognitive Impairment by Using 18F-FDG-PET: A Multicenter Study. Journal of Alzheimer’s Disease, 45(2), 543–552. https://doi.org/10.3233/JAD-141338 Jaisani, Z., Miletich, R. S., Ramanathan, M., & Weinstock, A. L. (2020). Clinical FDG-PET Findings in Patients with Temporal Lobe Epilepsy: Concordance with EEG and MRI. Journal of Neuroimaging: Official Journal of the American Society of Neuroimaging, 30(1), 119–125. https://doi.org/10.1111/jon.12671 Jakoby, B. W., Bercier, Y., Conti, M., Casey, M. E., Bendriem, B., & Townsend, D. W. (2011). Physical and clinical performance of the mCT time-of-flight PET/CT scanner. Physics in Medicine and Biology, 56(8), 2375– 2389. https://doi.org/10.1088/0031-9155/56/8/004 Jan, S., Santin, G., Strul, D., Staelens, S., Assié, K., Autret, D., Avner, S., Barbier, R., Bardiès, M., Bloomfield, P. M., Brasse, D., Breton, V., Bruyndonckx, P., Buvat, I., Chatziioannou, A. F., Choi, Y., Chung, Y. H., Comtat, C., Donnarieix, D., … Morel, C. (2004). GATE - Geant4 Application for Tomographic Emission: A simulation toolkit for PET and SPECT. Physics in Medicine and Biology, 49(19), 4543–4561. Jenkinson, M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M. W., & Smith, S. M. (2012). FSL. NeuroImage, 62(2), 782–790. https://doi.org/10.1016/j.neuroimage.2011.09.015 Kak, A. C., & Slaney, M. (2001). 3. Algorithms for Reconstruction with Nondiffracting Sources. In Principles of Computerized Tomographic Imaging (Vol. 1–0, pp. 49–112). Society for Industrial and Applied Mathematics. https://doi.org/10.1137/1.9780898719277.ch3 Karakatsanis, N. A., Lodge, M. A., Rahmim, A., & Zaidi, H. (2016). Introducing time-of-flight and resolution recovery image reconstruction to clinical whole-body PET parametric imaging. 2014 IEEE Nuclear Science Symposium and Medical Imaging Conference, NSS/MIC 2014, 7430771. https://doi.org/10.1109/NSSMIC.2014.7430771 Kato, T., Inui, Y., Nakamura, A., & Ito, K. (2016). Brain fluorodeoxyglucose (FDG) PET in dementia. Ageing Research Reviews, 30, 73–84. https://doi.org/10.1016/j.arr.2016.02.003 Keller, S. S., & Roberts, N. (2008). Voxel-based morphometry of temporal lobe epilepsy: An introduction and review of the literature. Epilepsia, 49(5), 741–757. https://doi.org/10.1111/j.1528-1167.2007.01485.x Kikuchi, K., Togao, O., Yamashita, K., Momosaka, D., Nakayama, T., Kitamura, Y., Kikuchi, Y., Baba, S., Sagiyama, K., Ishimatsu, K., Kamei, R., Mukae, N., Iihara, K., Suzuki, S. O., Iwaki, T., & Hiwatashi, A. (2021). Diagnostic accuracy for the epileptogenic zone detection in focal epilepsy could be higher in FDGPET/MRI than in FDG-PET/CT. European Radiology, 31(5), 2915–2922. https://doi.org/10.1007/s00330-020-07389-1 Kim, Y. K., Lee, D. S., Lee, S. K., Chung, C. K., Chung, J.-K., & Lee, M. C. (2002). 18F-FDG PET in Localization of Frontal Lobe Epilepsy: Comparison of Visual and SPM Analysis. Journal of Nuclear Medicine, 43(9), 1167–1174. Kinahan, P. E., Defrise, M., & Clackdoyle, R. (2004). CHAPTER 20—Analytic Image Reconstruction Methods. In M. N. Wernick & J. N. Aarsvold (Eds.), Emission Tomography (pp. 421–442). Academic Press. https://doi.org/10.1016/B978-012744482-6.50023-5 Klunk, W. E., Koeppe, R. A., Price, J. C., Benzinger, T. L., Devous, M. D., Jagust, W. J., Johnson, K. A., Mathis, C. A., Minhas, D., Pontecorvo, M. J., Rowe, C. C., Skovronsky, D. M., & Mintun, M. A. (2015). The Centiloid Project: Standardizing quantitative amyloid plaque estimation by PET. Alzheimer’s & Dementia: The Journal of the Alzheimer’s Association, 11(1), 1-15.e1-4. https://doi.org/10.1016/j.jalz.2014.07.003 Knesaurek, K., Warnock, G., Sideras, P., & Burger, C. (2014). Comparison of SUVr calculations in amyloid PET brain imaging. Journal of Nuclear Medicine, 55(supplement 1), 1873–1873. Knoll, G. F. (2000). Radiation detection and measurement (3rd ed). Wiley. Kochebina, O., Jan, S., Stute, S., Sharyy, V., Verrecchia, P., Mancardi, X., & Yvon, D. (2019). Performance Estimation for the High Resolution CaLIPSO Brain PET Scanner: A Simulation Study. IEEE Transactions on Radiation and Plasma Medical Sciences, 3(3), 363–370. https://doi.org/10.1109/TRPMS.2018.2880811 Kontaxakis, G. (2002). Iterative Image Reconstruction for Clinical PET Using Ordered Subsets, Median Root Prior, and a Web-Based Interface. Molecular Imaging & Biology, 4(3), 219–231. https://doi.org/10.1016/S1536-1632(02)00004-5 Kumar, A., & Chugani, H. T. (2017). The Role of Radionuclide Imaging in Epilepsy, Part 1: Sporadic Temporal and Extratemporal Lobe Epilepsy. Journal of Nuclear Medicine Technology, 45(1), 14–21. https://doi.org/10.2967/jnumed.112.114397 Kumar, A., Semah, F., Chugani, H. T., & Theodore, W. H. (2012). Chapter 26 - Epilepsy diagnosis: Positron emission tomography. In H. Stefan & W. H. Theodore (Eds.), Handbook of Clinical Neurology (Vol. 107, pp. 409–424). Elsevier. https://doi.org/10.1016/B978-0-444-52898-8.00026-4 Kurth, F., Luders, E., & Gaser, C. (2015). Voxel-Based Morphometry. In Brain Mapping (pp. 345–349). Elsevier. https://doi.org/10.1016/B978-0-12-397025-1.00304-3 Kuzniecky, R., de la Sayette, V., Ethier, R., Melanson, D., Andermann, F., Berkovic, S., Robitaille, Y., Olivier, A., Peters, T., & Feindel, W. (1987). Magnetic resonance imaging in temporal lobe epilepsy: Pathological correlations. Annals of Neurology, 22(3), 341–347. https://doi.org/10.1002/ana.410220310 Kuzniecky, R. I. (2005). Neuroimaging of Epilepsy: Therapeutic Implications. NeuroRx, 2(2), 384–393. Labbe, C., Thielemans, K., Belluzzo, D., Bettinardi, V., Gilardi, M., Hague, D., Jacobson, M., Kaiser, S., Levkovitz, R., Margalit, T., Mitra, G., Morel, C., Spinks, T., Valente, P., Zaidi, H., & Zverovich, A. (1999). An ObjectOriented Library for 3D PET Reconstruction Using Parallel Computing. https://doi.org/10.1007/978-3642-60125-5_50 Labbe, C., Thielemans, K., Zaidi, H., & Morel, C. (1999). An object-oriented library incorporating efficient projection/backprojection operators for volume reconstruction in 3D PET. Ladrón De Guevara, D. (2013). Rol del PET/CT en epilepsia. Revista Médica Clínica Las Condes, 24(6), 973–977. https://doi.org/10.1016/S0716-8640(13)70251-7 Landau, S. M., Fero, A., Baker, S. L., Koeppe, R., Mintun, M., Chen, K., Reiman, E. M., & Jagust, W. J. (2015). Measurement of Longitudinal β-Amyloid Change with 18F-Florbetapir PET and Standardized Uptake Value Ratios. Journal of Nuclear Medicine : Official Publication, Society of Nuclear Medicine, 56(4), 567–574. https://doi.org/10.2967/jnumed.114.148981 Lecoq, P., & Gundacker, S. (2021). SiPM applications in positron emission tomography: Toward ultimate PET time-of-flight resolution. The European Physical Journal Plus, 136(3), 292. https://doi.org/10.1140/epjp/s13360-021-01183-8 Lee, D. H., Gao, F.-Q., Rogers, J. M., Gulka, I., Mackenzie, I. R. A., Parrent, A. G., Kubu, C. S., Munoz, D. G., McLachlan, R. S., Blume, W. T., & Girvin, J. P. (1998). MR in Temporal Lobe Epilepsy: Analysis with Pathologic Confirmation. AJNR. American Journal of Neuroradiology, 19(1), 9. Lee, S. H., Bachman, A. H., Yu, D., Lim, J., & Ardekani, B. A. (2016). Predicting progression from mild cognitive impairment to Alzheimer’s disease using longitudinal callosal atrophy. Alzheimer’s & Dementia: Diagnosis, Assessment & Disease Monitoring, 2, 68–74. https://doi.org/10.1016/j.dadm.2016.01.003 Levin, C. S., Dahlbom, M., & Hoffman, E. J. (1995). A Monte Carlo correction for the effect of Compton scattering in 3-D PET brain imaging. IEEE Transactions on Nuclear Science, 42(4), 1181–1185. https://doi.org/10.1109/23.467880 Lewellen, T. K., Harrison, R. L., & Vannoy, S. (1998). The SimSET program, in Monte Carlo Calculations in Nuclear Medicine: Applications in Diagnostic Imaging. Institute of Physics Publication, Bristol, UK. Lorenzo, N. Y., Parisi, J. E., Cascino, G. D., Jack, C. R., Marsh, W. R., & Hirschorn, K. A. (1995). Intractable frontal lobe epilepsy: Pathological and MRI features. Epilepsy Research, 20(2), 171–178. https://doi.org/10.1016/0920-1211(94)00072-5 Lu, L., Zhang, H., Bian, Z., Ma, J., Feng, Q., & Chen, W. (2016). Validation of a Monte Carlo simulation of the Inveon PET scanner using GATE. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 828, 170–175. https://doi.org/10.1016/j.nima.2016.04.059 Maisey, M. N. (2005). Positron Emission Tomography in Clinical Medicine. In D. L. Bailey, D. W. Townsend, P. E. Valk, & M. N. Maisey (Eds.), Positron Emission Tomography: Basic Sciences (pp. 1–12). Springer. https://doi.org/10.1007/1-84628-007-9_1 Marshall, R. A., & Cully, C. M. (2020). Atmospheric effects and signatures of high-energy electron precipitation. In The Dynamic Loss of Earth’s Radiation Belts (pp. 199–255). Elsevier. https://doi.org/10.1016/B9780-12-813371-2.00007-X Marti-Fuster, B., Esteban, O., Thielemans, K., Setoain, X., Santos, A., Ros, D., & Pavia, J. (2014). Including anatomical and functional information in MC simulation of PET and SPECT brain studies. Brain-VISET: A voxel-based iterative method. IEEE Transactions on Medical Imaging, 33(10), 1931–1938. https://doi.org/10.1109/TMI.2014.2326041 Matsuda, H., Ito, K., Ishii, K., Shimosegawa, E., Okazawa, H., Mishina, M., Mizumura, S., Ishii, K., Okita, K., Shigemoto, Y., Kato, T., Takenaka, A., Kaida, H., Hanaoka, K., Matsunaga, K., Hatazawa, J., Ikawa, M., Tsujikawa, T., Morooka, M., … Sato, N. (2021). Quantitative Evaluation of 18F-Flutemetamol PET in Patients With Cognitive Impairment and Suspected Alzheimer’s Disease: A Multicenter Study. Frontiers in Neurology, 11, 578753. https://doi.org/10.3389/fneur.2020.578753 Matsuda, H., Yokoyama, K., Sato, N., Ito, K., Nemoto, K., Oba, H., Hanyu, H., Kanetaka, H., Mizumura, S., Kitamura, S., Shinotoh, H., Shimada, H., Suhara, T., Terada, H., Nakatsuka, T., Kawakatsu, S., Hayashi, H., Asada, T., Ono, T., … Shigemori, K. (2019). Differentiation Between Dementia With Lewy Bodies And Alzheimer’s Disease Using Voxel-Based Morphometry Of Structural MRI: A Multicenter Study. Neuropsychiatric Disease and Treatment, 15, 2715–2722. https://doi.org/10.2147/NDT.S222966 Mayoral, M., Marti‐Fuster, B., Carreño, M., Carrasco, J. L., Bargalló, N., Donaire, A., Rumià, J., Perissinotti, A., Lomeña, F., Pintor, L., Boget, T., & Setoain, X. (2016). Seizure-onset zone localization by statistical parametric mapping in visually normal 18F-FDG PET studies. Epilepsia, 57(8), 1236–1244. https://doi.org/10.1111/epi.13427 Mayoral, M., Niñerola-Baizán, A., Marti-Fuster, B., Donaire, A., Perissinotti, A., Rumià, J., Bargalló, N., SalaLlonch, R., Pavia, J., Ros, D., Carreño, M., Pons, F., & Setoain, X. (2019). Epileptogenic Zone Localization With 18FDG PET Using a New Dynamic Parametric Analysis. Frontiers in Neurology, 10. https://doi.org/10.3389/fneur.2019.00380 Mazzillo, M., Fallica, G., Ficarra, E., Messina, A., Romeo, M., & Zafalon, R. (2011). Solid state photodetectors for nuclear medical imaging applications. In Proceedings -Design, Automation and Test in Europe, DATE (p. 512). https://doi.org/10.1109/DATE.2011.5763091 Meikle, S. R., & Badawi, R. D. (2005). Quantitative Techniques in PET. In D. L. Bailey, D. W. Townsend, P. E. Valk, & M. N. Maisey (Eds.), Positron Emission Tomography: Basic Sciences (pp. 93–126). Springer. https://doi.org/10.1007/1-84628-007-9_5 Meltzer, C. C., Adelson, P. D., Brenner, R. P., Crumrine, P. K., Van Cott, A., Schiff, D. P., Townsend, D. W., & Scheuer, M. L. (2000). Planned ictal FDG PET imaging for localization of extratemporal epileptic foci. Epilepsia, 41(2), 193–200. https://doi.org/10.1111/j.1528-1157.2000.tb00139.x Merlin, T., Stute, S., Benoit, D., Bert, J., Carlier, T., Comtat, C., Filipovic, M., Lamare, F., & Visvikis, D. (2018). CASToR: A generic data organization and processing code framework for multi-modal and multidimensional tomographic reconstruction. Physics in Medicine and Biology, 63. https://doi.org/10.1088/1361-6560/aadac1 Meyer, P. T., Frings, L., Rücker, G., & Hellwig, S. (2017). 18F-FDG PET in Parkinsonism: Differential Diagnosis and Evaluation of Cognitive Impairment. Journal of Nuclear Medicine, 58(12), 1888–1898. https://doi.org/10.2967/jnumed.116.186403 Monnier, F., Fayad, H., Bert, J., Schmidt, H., & Visvikis, D. (2015). Validation of a simultaneous PET/MR system model for PET simulation using GATE. EJNMMI Physics, 2(1), A45. https://doi.org/10.1186/2197-73642-S1-A45 Mosconi, L., De Santi, S., Li, Y., Li, J., Zhan, J., Tsui, W. H., Boppana, M., Pupi, A., & de Leon, M. J. (2006). Visual rating of medial temporal lobe metabolism in mild cognitive impairment and Alzheimer’s disease using FDG-PET. European Journal of Nuclear Medicine and Molecular Imaging, 33(2), 210–221. https://doi.org/10.1007/s00259-005-1956-z Moses, W. W. (2011). Fundamental Limits of Spatial Resolution in PET. Nuclear Instruments & Methods in Physics Research. Section A, Accelerators, Spectrometers, Detectors and Associated Equipment, 648 Supplement 1, S236–S240. https://doi.org/10.1016/j.nima.2010.11.092 Mota, A., Cuplov, V., Schott, J., Hutton, B., Thielemans, K., Drobnjak, I., Dickson, J., Bert, J., Burgos, N., Cardoso, J., Modat, M., Ourselin, S., & Erlandsson, K. (2015). Establishment of an open database of realistic simulated data for evaluation of partial volume correction techniques in brain PET/MR. EJNMMI Physics, 2(1), A44. https://doi.org/10.1186/2197-7364-2-S1-A44 Mountz, J. M., Laymon, C. M., Cohen, A. D., Zhang, Z., Price, J. C., Boudhar, S., McDade, E., Aizenstein, H. J., Klunk, W. E., & Mathis, C. A. (2015). Comparison of qualitative and quantitative imaging characteristics of [11C]PiB and [18F]flutemetamol in normal control and Alzheimer’s subjects. NeuroImage. Clinical, 9, 592–598. https://doi.org/10.1016/j.nicl.2015.10.007 Mueller, S. G., Laxer, K. D., Cashdollar, N., Buckley, S., Paul, C., & Weiner, M. W. (2006). Voxel-based Optimized Morphometry (VBM) of Gray and White Matter in Temporal Lobe Epilepsy (TLE) with and without Mesial Temporal Sclerosis. Epilepsia, 47(5), 900–907. https://doi.org/10.1111/j.15281167.2006.00512.x Muhlhofer, W., Tan, Y.-L., Mueller, S. G., & Knowlton, R. (2017). MRI-negative temporal lobe epilepsy—What do we know? Epilepsia, 58(5), 727–742. https://doi.org/10.1111/epi.13699 National Electrical Manufacturers Association. (1994). Standards Publication NU-2-1994: Performance measurements of positron emission tomography. Washington DC: National Electrical Manufacturers Association. National Electrical Manufacturers Association. (2007). NEMA Standards Publication NU 2-2007: Performance Measurements of Positron Emission Tomographs. National Electrical Manufacturers Association. https://books.google.es/books?id=WhutMwEACAAJ Niñerola-Baizán, A., Aguiar, P., Cabrera-Martín, M. N., Vigil, C., Gómez-Grande, A., Lorenzo, C., Rubí, S., Sopena, P., Camacho, V., & por el Grupo de Neuroimagen SEMNIM. (2020). Relevance of quantification in brain PET studies with 18F-FDG. Revista Espanola De Medicina Nuclear E Imagen Molecular, 39(3), 184–192. https://doi.org/10.1016/j.remn.2020.03.003 Nooraine, J., Iyer, R. B., & Raghavendra, S. (2013). Ictal PET in presurgical workup of refractory extratemporal epilepsy. Annals of Indian Academy of Neurology, 16(4), 676–677. https://doi.org/10.4103/09722327.120475 Ollinger, J. M. (1996). Model-based scatter correction for fully 3D PET. Physics in Medicine & Biology, 41(1), 153. https://doi.org/10.1088/0031-9155/41/1/012 Ortner, M. M. (2018). The Use of 18F-FDG PET in the Diagnostic Workup of Alzheimer’s Dementia. In R. Perneczky (Ed.), Biomarkers for Alzheimer’s Disease Drug Development (pp. 213–219). Springer. https://doi.org/10.1007/978-1-4939-7704-8_14 Pagani, M., Nobili, F., Morbelli, S., Arnaldi, D., Giuliani, A., Öberg, J., Girtler, N., Brugnolo, A., Picco, A., Bauckneht, M., Piva, R., Chincarini, A., Sambuceti, G., Jonsson, C., & De Carli, F. (2017). Early identification of MCI converting to AD: A FDG PET study. European Journal of Nuclear Medicine and Molecular Imaging, 44(12), 2042–2052. https://doi.org/10.1007/s00259-017-3761-x Whitwell, J. L. (2009). Voxel-Based Morphometry: An Automated Technique for Assessing Structural Changes in the Brain. The Journal of Neuroscience, 29(31), 9661–9664. https://doi.org/10.1523/JNEUROSCI.216009.2009 Whitwell, J. L., Shiung, M. M., Przybelski, S., Weigand, S. D., Knopman, D. S., Boeve, B. F., Petersen, R. C., & Jack, C. R. (2008). MRI patterns of atrophy associated with progression to AD in amnestic Mild Cognitive Impairment. Neurology, 70(7), 512–520. https://doi.org/10.1212/01.wnl.0000280575.77437.a2 Willmann, O., Wennberg, R., May, T., Woermann, F. G., & Pohlmann-Eden, B. (2007). The contribution of 18FFDG PET in preoperative epilepsy surgery evaluation for patients with temporal lobe epilepsy. Seizure, 16(6), 509–520. https://doi.org/10.1016/j.seizure.2007.04.001 Winston, G., Micallef, C., Kendell, B., Bartlett, P., Williams, E., Burdett, J., & Duncan, J. (2013). The value of repeat neuroimaging for epilepsy at a tertiary referral centre: 16 years of experience. Epilepsy Research, 105. https://doi.org/10.1016/j.eplepsyres.2013.02.022 Wong, D. F., Links, J. M., Molliver, M. E., Hengst, T. C., Clifford, C. M., Buhle, L., Bryan, M., Stumpf, M., & Wagner, H. N. Jr. (1984). An anatomically realistic brain phantom for quantification with positron tomography. Journal of Nuclear Medicine, 25(5), 108. Yamane, T., Ishii, K., Sakata, M., Ikari, Y., Nishio, T., Ishii, K., Kato, T., Ito, K., Senda, M., & J-ADNI Study Group. (2017). Inter-rater variability of visual interpretation and comparison with quantitative evaluation of 11C-PiB PET amyloid images of the Japanese Alzheimer’s Disease Neuroimaging Initiative (J-ADNI) multicenter study. European Journal of Nuclear Medicine and Molecular Imaging, 44(5), 850–857. https://doi.org/10.1007/s00259-016-3591-2 Yao, Z., Hu, B., Zhao, L., & Liang, C. (2011). Analysis of Gray Matter in AD Patients and MCI Subjects Based Voxel-Based Morphometry. In B. Hu, J. Liu, L. Chen, & N. Zhong (Eds.), Brain Informatics (pp. 209–217). Springer. https://doi.org/10.1007/978-3-642-23605-1_22 Yasuda, C. L., Betting, L. E., & Cendes, F. (2010). Voxel-based morphometry and epilepsy. Expert Review of Neurotherapeutics, 10(6), 975–984. https://doi.org/10.1586/ern.10.63 Zaidi, H. (1999). Relevance of accurate Monte Carlo modeling in nuclear medical imaging. Medical Physics, 26(4), 574–608. https://doi.org/10.1118/1.598559 Zaidi, H., Montandon, M.-L., & Alavi, A. (2007). Advances in Attenuation Correction Techniques in PET. PET Clinics, 2(2), 191–217. https://doi.org/10.1016/j.cpet.2007.12.002 Zanzonico, P. (2004). Positron emission tomography: A review of basic principles, scanner design and performance, and current systems. Seminars in Nuclear Medicine. https://doi.org/10.1053/J.SEMNUCLMED.2003.12.002 Zvolský, M. (2014, November 28). Tomographic Image Reconstruction. - An Introduction. https://www.desy.de/~garutti/LECTURES/BioMedical/Lecture7_ImageReconstruction.pdf Zvolský, M., Horns, D., & Garutti, E. (2017). Simulation, Image Reconstruction and SiPM Characterisation for a Novel Endoscopic Positron Emission Tomography Detector [Universität Hamburg]. https://bibpubdb1.desy.de/record/398074/files/PhDThesis_published.pdf APPENDICES 7 APPENDICES The content of this chapter has already been published as Paredes-Pacheco, J. et al. SimPETAn open online platform for the Monte Carlo simulation of realistic brain PET data. Validation for 18 F-FDG scans. Med Phys. 2021 May; 48(5):2482-2493. doi: 10.1002/mp.14838. PMID: 33713354; PMCID: PMC8252452. 7.1 SIMULATION OF PATHOLOGICAL MAPS SimPET can be used to validate quantitative methods in different pathologies, in addition to cortical epilepsy. Users can download the generated activity maps, and modify them to introduce known hypometabolism patterns, which can be used as ground-truth for the quantification after the simulation and reconstruction. As an example, one of the generated 18FFDG activity maps was downloaded, and two different hypometabolism were added manually in predefined ROIs reflecting patterns typically observed in AD and TLE. The simulations of this patterns in the three scanners are showcased in Figure A1, allowing for the different representation of the exact same metabolism in the different scanners to be easily observed. Figure A1: Pipeline used to simulate pathological PET images of AD and TLE cases using activity maps generated on SimPET and including a known level of hypometabolism. After simulating the generated patterns using the platform, measurements on the different scanners can be compared, allowing the users to validate quantification and harmonization protocols. 7.2 AMYLOID-PET SIMULATION In addition to 18F-FDG, SimPET also allows users to simulate amyloid-PET images with 18FFlorbetapir as a tracer. For demonstrating this capability, we include a simple example case. Five amyloid-negative patients, acquired on the GE Discovery ST, were uploaded to the platform to obtain amyloid-PET realistic activity maps. These maps were then simulated on the three implemented scanners and the results are showcased for demonstration purposes, showing the differences between scanners in amyloid-PET assessments. Figure A2: Sample of five amyloid-negative patient PET images acquired on GE Discovery ST, the corresponding generated activity maps, and the simulated PET images obtained in the different commercial scanners, GE Advance NXi and Siemens Biograph mCT images are included for comparison. Figure A2 shows the results of the performed amyloid-PET simulations. Five amyloid-negative PET patients and the obtained activity maps are shown in rows 1 and 2, followed by the simulated PETs obtained with the platform for the GE Discovery ST, GE Advance NXi and Siemens Biograph mCT scanner models (rows 3-5). Real amyloid-negative PET examples for the GE Advance NXi and Siemens Biograph mCT scanners are included for comparison. DECLARATIONS 8 DECLARATIONS: USE OF IMAGES, TABLES AND PUBLISHED CONTENT 8.1 JOURNAL AUTHORIZATIONS FOR THE USE OF IMAGES AND TABLES  Table 1 and Figure 12  Figure 2 and Table 2