Diffusion magnetic resonance imaging (dMRI) allows uniquely the study of the human brain non-invasively and in vivo. Advances in dMRI offer new insight into tissue microstructure and connectivity, and the possibility of investigating the mechanisms and pathology of neurological diseases. The great potential of the technique relies on indirect inference, as modelling frameworks are necessary to map dMRI measurements to neuroanatomical features. However, this mapping can be computationally expensive, particularly given the trend of increasing dataset sizes and/or the increased complexity in biophysical modelling. Limitations on computing can restrict data exploration and even methodology development. A step forward is to take advantage of the power offered by recent parallel computing architectures, especially Graphics Processing Units (GPUs). GPUs are massive parallel processors that offer trillions of floating point operations per second, and have made possible the solution of computationally intensive scientific problems that were intractable before. However, they are not inherently suited for all types of problems, and bespoke computational frameworks need to be developed in many cases to take advantage of their full potential. In this thesis, we propose parallel computational frameworks for the analysis of dMRI using GPUs within different contexts. We show that GPU-based designs can offer accelerations of more than two orders of magnitude for a number of scientific computing tasks with different parallelisability requirements, ranging from biophysical modelling for tissue microstructure estimation to white matter tractography for connectome generation. We develop novel and efficient GPUaccelerated solutions, including a framework that automatically generates GPU parallel code from a user-specified biophysical model. We also present a parallel GPU framework for performing probabilistic tractography and generating whole-brain connectomes. Throughout the thesis, we discuss several strategies for parallelising scientific applications, and we show the great potential of the accelerations obtained, which change the perspective of what is computationally feasible.

Une nouvelle stratégie oncologique, basée sur l’intégration de la radiothérapie nanovectorisée et l’administration loco-régionale, a été évaluée pour le traitement et l’imagerie du glioblastome, le type le plus commun des tumeurs cérébrales primaires. Les dendrimères Gallic Acid-Triethylène Glycol (GATG) sont des nanovecteurs de choix pour délivrer simultanément l’agent thérapeutique (le radioisotope 188Re par son rayonnement béta a été retenu) et l’agent diagnostique (le gadolinium est un agent paramagnétique utilisé en Imagerie par Résonance Magnétique (IRM)). Leur évaluation a été réalisée par administration locorégionale par stéréotaxie sur un modèle de rat F 98. Les données pharmaco-cinétiques ont été également obtenues après injection intraveineuse permettant d’apprécier les propriétés des différents dendrimères synthétisés. Leur apport en terme de confinement au site d’injection représente un avantage majeur de ce nouveau type de radiopharmaceutiques. / A new oncologic strategy, based on the integration of nanovectorized radiotherapy and locoregional delivery, was evaluated for the treatment and imaging of glioblastomas, the most common and lethal type of primary brain tumors. Gallic acidtriethylene glycol (GATG) dendrimers were the nanovectors of choice to deliver the radiotherapeutic 188Re and paramagnetic nuclei Gd3+, with a minimally invasive stereotactic injection, directly depositing the radiotherapeutic dose to the tumor site in a F98 rat glioma model. Intravenous injection was used to further investigate the pharmacokinetics, throughout body distribution and clearance profiles of these dendrimers. Molecular weight and architecture had an important role on the in vivo behavior of the dendrimers. Their use as nanovectors prevented the fast brain clearance of the radionuclide alone, and prolonged the confinement of the internal radiation at the tumor site.

Bien qu'outil de référence pour le diagnostic et le suivi de la SEP, l'IRM conventionnelle ne reste que modérément corrélée à l'état clinique du patient. Afin de mieux caractériser les altérations pathologiques, nous employons dans ce travail les techniques d'IRM dites non conventionnelles que sont la spectroscopie par résonance magnétique (SRM) et l'IRM de diffusion. Un premier suivi hebdomadaire, a permis de mettre en évidence la sensibilité des métriques de diffusion et la spécificité de la SRM pour détecter les processus initiaux de la formation d'une lésion.Un second suivi a permis de mettre en évidence des modifications de la diffusivité dans plusieurs faisceaux de substance blanche, avec notamment une diminution de la fraction d'anisotropie et une augmentation de diffusivité radiale, s'aggravant avec l'avancée de la maladie et plus marquée dans les formes progressives.Enfin, l'application de la théorie des graphes a permis de caractériser la connectivité cérébrale dans les quatre formes cliniques et d'étudier leur évolution. Cette étude a permis de mettre en évidence des altérations dans tous les phénotypes cliniques, avec notamment une diminution de la densité du réseau cérébral, plus importante dans les formes progressives de la maladie et tendant à s'accentuer avec la progression de la maladie.Ce travail montre la sensibilité des techniques avancées d'IRM pour la caractérisation des altérations pathologiques et de leur évolution dans la SEP / While conventional MRI is the reference tool for the diagnosis and monitoring of MS, it remains only moderately correlated with the patient’s clinical status. In order to better characterize pathological alterations occurring in MS, we use in this work non-conventional MRI techniques, namely magnetic resonance spectroscopy (MRS) and diffusion MRI.A first weekly follow-up revealed the sensitivity of the diffusion metrics and the specificity of the SRM to detect the initial processes of lesion formation.A second follow-up revealed changes in diffusivity in several white matter fiber bundles, including a decrease in fraction of anisotropy and an increase in radial diffusivity, worsening with advancing disease and more marked in the progressive forms.Finally, the application of graph theory allowed to characterize the brain connectivity in the four clinical forms and to study their evolution. This study allowed us to highlight alterations in all the four clinical phenotypes, including a decrease in the cerebral network density, more marked in the progressive forms of the disease and tending to increase with its progression.This work shows the sensitivity of advanced MRI techniques for the characterization of pathological alterations and their evolution in MS

Glioblastoma multiform (GBM) is an aggressive primary brain tumour. Despite treatment advances in recent years, outcomes remain poor. Disease progression tends to occur adjacent to the original tumour or surgical resection bed, usually within the radiotherapy planning field. This local recurrence and progression is believed to be the result of invasive disease in the surrounding tissue at the time of diagnosis and treatment, which is not currently detectable by conventional non-invasive methods. A number of novel therapies are currently under development which target specific aspects of the tumour behaviour, to try and improve outcomes from this devastating disease. Imaging biomarkers are under development, therefore, in order to provide a non-invasive assessment of tumour extent and behaviour, to provide bespoke image-guided therapies, and detect recurrence or treatment failure at the earliest opportunity. These are also of value in the context of novel therapeutics, which may have a very specific affect on an aspect of tumour behaviour that is not readily apparent on standard clinical imaging. Key to the progression of GBM is the invasion into surrounding white matter. This is followed by a period of tumour growth and subsequent angiogenesis in which microvasculature is produce that is distinct from the highly regulated blood-brain barrier. This thesis covers the development of suite of advanced magnetic resonance imaging (MRI) techniques aimed at characterising those very traits of GBM responsible for the aggressiveness and treatment resistance. Repeatability studies are undertaken to determine the performance of the biomarkers in healthy tissues, and also in a range of gliomas. Dynamic Contrast Enhanced (DCE-) and dynamic susceptibility-enhanced (DSC-)MRI are used to provide estimates of perfusion and permeability in the tumour. In order to address the reasons behind preferential invasion of the corpus callosum, they are used in conjunction with ASL to non-invasively map perfusion territories and watershed regions in the brain through perfusion timing parameters. Diffusion Tensor Imaging (DTI) and quantitative magnetisation transfer (qMT) are used to provide complementary information about white matter integrity, in order to identify changes occurring with glioma invasion as early as possible and direct image-guided treatments at presentation. Their complementary nature is assessed by comparing the two parameters simultaneously in white matter. Additionally, one of the qMT parameters which may be related to tissue pH is shown to be sensitive and specific for the detection of high-grade tumour tissue. Finally, a novel multiparametric imaging biomarker is introduced. Tumour surface mapping assesses the boundary between the solid tumour and surrounding tissue in order to identify areas of potential aggressiveness and invasion. Multiple imaging parameters can be combined to improve specificity and sensitivity. Using the diffusion-weighted imaging parameter, mean diffusivity (MD - also referred to as the apparent diffusion coefficient (ADC)), it is shown to be predictive of clinical outcome in a retrospective and prospective study, while a multiparametric example is given indicating the utility as a predicative biomarker for regions of progression and recurrence, and as potential spatial discriminator for image-guided therapies.

Les techniques d'imagerie par résonance magnétique de Diffusion (IRMd) et de Perfusion (IRMp) permettent la détection de divers aspects importants et complémentaires en imagerie cérébrale. Le travail effectué dans cette thèse présente des contributions théoriques et méthodologiques sur les modalités d'IRM basées sur des images pondérées en diffusion, et sur des images de perfusion par injection de produit de contraste. Pour chacune des deux modalités, les contributions de la thèse sont liées au développement de nouvelles méthodes pour améliorer la qualité, le traitement et l'exploitation des signaux acquis. En IRM de diffusion, la nature complexe du signal est étudiée avec un accent sur l'information de phase. Le signal complexe est ensuite exploité pour corriger le biais induit par le bruit d'acquisition des images, améliorant ainsi l'estimation de certaines métriques structurelles. En IRM de perfusion, le traitement du signal est revisité afin de tenir compte du biais dû à la dispersion du bolus. On montre comment ce phénomène, qui peut empêcher la correcte estimation des métriques de perfusion, peut aussi donner des informations importantes sur l'état pathologique du tissu cérébral. Les contributions apportées dans cette thèse sont présentées dans un cadre théorique et méthodologique validé sur de nombreuses données synthétiques et réelles. / Diffusion and Perfusion Magnetic Resonance Imaging (dMRI & pMRI) represent two modalities that allow sensing important and different but complementary aspects of brain imaging. This thesis presents a theoretical and methodological investigation on the MRI modalities based on diffusion-weighted (DW) and dynamic susceptibility contrast (DSC) images. For both modalities, the contributions of the thesis are related to the development of new methods to improve and better exploit the quality of the obtained signals. With respect to contributions in diffusion MRI, the nature of the complex DW signal is investigated to explore a new potential contrast related to tissue microstructure. In addition, the complex signal is exploited to correct a bias induced by acquisition noise of DW images, thus improving the estimation of structural scalar metrics. With respect to contributions in perfusion MRI, the DSC signal processing is revisited in order to account for the bias due to bolus dispersion. This phenomenon prevents the correct estimation of perfusion metrics but, at the same time, can give important insights about the pathological condition of the brain tissue. The contributions of the thesis are presented within a theoretical and methodological framework, validated on both synthetical and real images.

Diffusion MRI (dMRI) measurements allow us to infer the microstructural properties of white matter and to reconstruct fiber pathways in-vivo. High angular diffusion imaging (HARDI) allows for the creation of more and more complex local models connecting the microstructure to the measured signal. One of the challenges is the derivation of meaningful metrics describing the underlying structure from the local models. The aim hereby is to increase the specificity of the widely used metric fractional anisotropy (FA) by using the additional information contained within the HARDI data. A local model which is connected directly to the underlying microstructure through the model of a single fiber population is spherical deconvolution. It produces a fiber orientation density function (fODF), which can often be interpreted as superposition of multiple peaks, each associated to one relatively coherent fiber population (bundle). Parameterizing these peaks one is able to disentangle and characterize these bundles. In this work, the fODF peaks are approximated by Bingham distributions, capturing first and second order statistics of the fiber orientations, from which metrics for the parametric quantification of fiber bundles are derived. Meaningful relationships between these measures and the underlying microstructural properties are proposed. The focus lies on metrics derived directly from properties of the Bingham distribution, such as peak length, peak direction, peak spread, integral over the peak, as well as a metric derived from the comparison of the largest peaks, which probes the complexity of the underlying microstructure. These metrics are compared to the conventionally used fractional anisotropy (FA) and it is shown how they may help to increase the specificity of the characterization of microstructural properties. Visualization of the micro-structural arrangement is another application of dMRI. This is done by using tractography to propagate the fiber layout, extracted from the local model, in each voxel. In practice most tractography algorithms use little of the additional information gained from HARDI based local models aside from the reconstructed fiber bundle directions. In this work an approach to tractography based on the Bingham parameterization of the fODF is introduced. For each of the fiber populations present in a voxel the diffusion signal and tensor are computed. Then tensor deflection tractography is performed. This allows incorporating the complete bundle information, performing local interpolation as well as using multiple directions per voxel for generating tracts. Another aspect of this work is the investigation of the spherical harmonic representation which is used most commonly for the fODF by means of the parameters derived from the Bingham distribution fit. Here a strong connection between the approximation errors in the spherical representation of the Dirac delta function and the distribution of crossing angles recovered from the fODF was discovered. The final aspect of this work is the application of the metrics derived from the Bingham fit to a number of fetal datasets for quantifying the brain’s development. This is done by introducing the Gini-coefficient as a metric describing the brain’s age.

Background. Neuroimaging studies have demonstrated an association between lithium (Li) treatment and brain structure in human subjects. A crucial unresolved question is whether this association reflects direct neurochemical effects of Li or indirect effects secondary to treatment or prevention of episodes of bipolar disorder (BD). Method. To address this knowledge gap, we compared manually traced hippocampal volumes in 37 BD patients with at least 2 years of Li treatment (Li group), 19 BD patients with <3 months of lifetime Li exposure over 2 years ago (non-Li group) and 50 healthy controls. All BD participants were followed prospectively and had at least 10 years of illness and a minimum of five episodes. We established illness course and long-term treatment response to Li using National Institute of Mental Health (NIMH) life charts. Results. The non-Li group had smaller hippocampal volumes than the controls or the Li group (F₂,₁₀₂ =4.97, p=0.009). However, the time spent in a mood episode on the current mood stabilizer was more than three times longer in the Li than in the non-Li group (t₅₁ =2.00, p=0.05). Even Li-treated patients with BD episodes while on Li had hippocampal volumes comparable to healthy controls and significantly larger than non-Li patients (t₄₃=2.62, corrected p=0.02). Conclusions. Our findings support the neuroprotective effects of Li. The association between Li treatment and hippocampal volume seems to be independent of long-term treatment response and occurred even in subjects with episodes of BD while on Li. Consequently, these effects of Li on brain structure may generalize to patients with neuropsychiatric illnesses other than BD.

Introduction Two current major research topics in nuclear medicine are in the development of long-lived positron-emitting nuclides for imaging tracers with long biological half-lives and in theranostics, imaging nuclides which have a chemically analogous therapy isotope. As shown in TABLE 1, the radioisotopes of arsenic (As) are well suited for both of these tasks with several imaging and therapy isotopes of a variety of biologically relevant half-lives accessible through the use of small medical cyclotrons. The five naturally abundant isotopes of germanium are both a boon and challenge for the medical nuclear chemist. They are beneficial in that they facilitate a wide array of producible radioarsenic isotopes. They are a challenge as monoisotopic radioarsenic production requires isotopically-enriched targets that are expensive and of limited availability. This makes it highly desirable that the germanium target material is reclaimed from arsenic isolation chemistry. One major factor which has limited the development of radioarsenic has been difficulties in its incorporation into biologically relevant targeting vectors. Previous studies have labeled antibodies and polymers through covalent bonding of arsenite (As(III)) with the sulfydryl group1,2,3. Recent work in our group has shown the facile synthesis and utility of superparamagnetic iron oxide nanoparticle- (SPION-)bound radioarsenic as a dual modality positron emission tomography (PET)/magnetic resonance imaging (MRI) agent4. Presently, we have built upon previous studies producing, isolating, and labeling untargeted SPION with radioarsenic4,5. We have incorp-rated the use of isotopically-enriched 72GeO2 for the production of radioisotopically pure 72As. The bulk of the 72GeO2 target material was re-claimed from the arsenic isolation chemical procedure for reuse in future irradiations. The 72As was used for ongoing development toward the synthesis of targeted, As-SPION-based, dual-modality PET/MRI agents. Material and Methods Targets of ~100 mg of isotopically-enriched 72GeO2 (96.6% 72Ge, 2.86% 73Ge, 0.35% 70Ge, 0.2% 74Ge, 0.01% 76Ge, Isoflex USA) were pressed into a niobium beam stop at 225 MPa, covered with a 25 µm HAVAR containment foil, attached to a water-cooling target port, and irradiated with 3 µA of 16.1 MeV protons for 2–3 hours using a GE PETtrace cyclotron. After irradiation, the target and beam stop were assembled into a PTFE dissolution apparatus, where the 72GeO2 target material was dissolved with the addition of 2 mL of 4 M NaOH and subsequent stirring. After dissolution was completed, the clear, colorless solution was transferred to a fritted glass column and the bulk 72GeO2 was reprecipitated by neutralizing the solution with the addition of 630 µL [HCl]conc, filtered, and rinsed with 1 mL [HCl]conc. To the combined 72As-containing filtrates, 100 µL 30% H2O2 was added to ensure that 72As was in the nonvolatile As(V) oxidation state. The ~3 mL solution was then evaporated at 115 ˚C while the vessel was purged with argon, followed by a second addition of 100 µL H2O2 after the volume was reduced to 1 mL. When the filtrate volume was ~0.3 mL, the vessel was removed from heat, allowed to cool with argon flow, and the arsenic reconstituted in 1 mL [HCl]conc and loaded onto a 1.5 mL bed volume Bio-Rad AG 1×8, 200–400 mesh anion exchange column preconditioned with 10 M HCl. The radioarsenic was eluted in 10 M HCl in the next ~10 mL, with 90% of the activity eluting in a 4 mL fraction. The column was then eluted with 5 mL 1 M HCl. The 72As-rich 10 M HCl fraction was reduced to As(III) with the addition of ~100 mg CuCl, and heating to 60 ˚C for 1 hour. The resulting AsCl3 was then extracted twice into 4 mL cyclohexane, which were combined and back extracted into 500 µL of water as As(OH)3. This solution of 72As in H2O was then used directly to label SPION and for subsequent experiments conjugating 72As-SPION with TRC105, an angiogenesis-marking monoclonal antibody (MAb) targeting endoglin/CD105. Several methods were initially attempted involving directly conjugating the surface-modified SPION to the MAb through a polyethylene glycol (PEG) linker. More recent studies have investigated the radioarsenic labeling of SPION encapsulated in hollow mesoporous silica nanoparticles (SPION@HMSN) and its subsequent conjugation to TRC105. Results and Conclusion Irradiation of pressed, isotopically-enriched 72GeO2 resulted in a production yield for 72As of 17 ± 2 mCi/(µA·hr·g) and for 71As of 0.37 ± 0.04 mCi/(µA·hr·g), which are 64 % and 33 %, of those predicted from literature6, respectively. However, these production yields are in agreement with those scaled from observed production yields using analagous natGeO2 targets. The end-of-bombardment 72As radionuclidic purity can be improved by minimizing the 72Ge(p,2n)71As reaction by degrading the beam energy. A 125 µm Nb containment foil would degrade impinging protons to 14.1 MeV and is predicted to reduce 71As yield by a factor of three, while only reducing 72As yield by 1 %6, improving end-of-bombardment radionuclidic purity from 98 % to greater than 99 %. Overall decay-corrected radiochemical yield of the 72As isolation procedure from 72GeO2 were 51 ± 2 % (n = 3) in agreement with those observed with natGeO2 57 ± 7 % (n = 14). The beam current was limited to 3 µA as higher cur-rents 4–5 µA exhibited inconsistent dissolution and reprecipitation steps, resulting in an overall yield of 44 ± 21 % (n = 6). Dissolution time also played an important role in overall yield with at least one hour necessary to minimize losses in these first two steps. The separation procedure effectively removed all radiochemical contaminants and resulted in 72As(OH)3 isolated in a small volume, pH~4.5 water solution. Over the course of minutes to hours after back extraction, rapid auto-oxidation to 72AsO4H3 was observed. The bulk 72GeO2 target material, which was reclaimed from the isolation procedure, is being collected for future use. The synthesis of a targeted PET/MRI agent based on the functionalization of 72As-SPION has proved to be a difficult task. Experiments conjugating 72As-SPION to TRC105 through a PEG linker were unsuccessful, despite the investigation of a variety bioconjugation procedures. Current work is investigating the use of SPION@HMSN, which have a similar affinity for 72As as unencapsulated SPION. This new class of 72As-labeled SPION@HMSN has a hollow cavity for potential anti-cancer drug loading, as well as the mesoporous silica surface, which may facilitate the efficient conjugation of TRC105 using a well-developed bioconjugation technique. In summary, radioarsenic holds potential in the field of diagnostic and therapeutic nuclear medicine. However, this potential remains locked behind challenges related to its production and useful in vivo targeting. The present work strives to address several of these challenges through the use of enriched 72GeO2 target material, a chemical isolation procedure that reclaims the bulk of the target material, and the investigation of new targeted nanoparticle-based PET/MRI agents.

Post-traumatic stress disorder (PTSD) is a heterogenous psychiatric condition that affects thousands of individuals each year. Of those who experience this condition, military members including members of the Canadian Armed Forces (CAF) are particularly vulnerable, demonstrating high prevalence rates of PTSD-related symptoms. Moreover, individuals with PTSD are at increased risk for comorbid conditions and are at greater risk for suicide due to the overwhelming, debilitating nature of PTSD symptoms. In previous research, hyperarousal symptoms associated with PTSD have been linked to dysregulation in the locus coeruleus norepinephrine (LC-NE) system, a vast neuromodulatory system responsible for regulating arousal, attention, autonomic and memory-related functions. Advancements in neuroimaging methods have advanced our ability to study connectivity in vivo such that small structures like the LC can be further studied in human samples. Specifically, neuromelanin-sensitive MRI (NM-MRI), a novel, non-invasive neuroimaging method has been shown to detect changes in neuromelanin (NM)-related signal in both the LC and substantia nigra (SN). NM is a dark pigment that accumulates over the lifespan in catecholamine-dominant centers such as the LC and SN and is the by-product of catecholamine oxidation. NM-MRI can be used to image these centers in vivo due to the paramagnetic properties offered by NM. Furthermore, when excess cytosolic catecholamine levels are present in select neurons, NM production is thought to be increased, resulting in increased NM signal from the LC. This could potentially be a marker for dysregulation as many conditions have been associated with variability of this system. Previously, NM-MRI has been used in other clinical settings such as in Parkinson’s disease (PD), Alzheimer’s disease (AD), schizophrenia and depression; however, this current investigation is the first to utilize this imaging modality in the context of PTSD. Specifically, we hypothesized that increased NM-MRI signal in the LC would correlate with increasing severity of hyperarousal symptoms in individuals with PTSD. We also predicted that the opposite would be true for comorbid depression symptom severity, as reduced LC signal has been previously correlated with clinical measures of comorbid depression using NM-MRI. As per our primary hypothesis, we observed a significant positive correlation between NM-MRI signals in the caudal elements of the LC with hyperarousal symptom severity in 22 PTSD subjects (r= 0.54, p= 0.017; partial correlation controlling for depression symptom severity, age, and sex). In contrast, we did not find any evidence to support our secondary hypothesis, because a non-significant trend correlating LC NM-MRI signal and depression symptom severity was obtained (r= -0.30, p=0.22; partial correlation controlling for hyperarousal severity, age, and sex). Based on these results, we were able to build on previously conducted work to further investigate the utility of NM-MRI in the detection of variability in LC-NE system as it pertains to psychiatric conditions known to show dysregulation of this system such as PTSD. In addition, this thesis provides further evidence to support the automation of NM-MRI analytical methods, thus supporting their potential utility for future clinical research. Our findings also provide support for the use of NM-MRI as a potential measure of NE activity; further, this work provided preliminary evidence supporting the use of NM-MRI in a clinical, psychiatric setting, where the technique may serve as a biomarker of PTSD pathology. With these findings in mind, additional validation studies can be conducted to verify the use of NM-MRI as a biomarker for NE system dysregulation. This would potentially allow for advancements in targeted treatment options for PTSD, particularly those targeting the LC-NE system, thus potentially increasing patient stratification and treatment efficacy.

Segmentation of magnetic resonance images is an important part of planning radiotherapy treat-ments for patients with brain tumours but due to the number of images contained within a scan and the level of detail required, manual segmentation is a time consuming task. Convolutional neural networks have been proposed as tools for automated segmentation and shown promising results. However, the data sets used for training these deep learning models are often imbalanced and contain data that does not contribute to the performance of the model. By carefully selecting which data to train on, there is potential to both speed up the training and increase the network’s ability to detect tumours. This thesis implements the method of importance sampling for training a convolutional neural network for patch-based segmentation of three dimensional multimodal magnetic resonance images of the brain and compares it with the standard way of sampling in terms of network performance and training time. Training is done for two different patch sizes. Features of the most frequently sampled volumes are also analysed. Importance sampling is found to speed up training in terms of number of epochs and also yield models with improved performance. Analysis of the sampling trends indicate that when patches are large, small tumours are somewhat frequently trained on, however more investigation is needed to confirm what features may influence the sampling frequency of a patch. / Segmentering av magnetröntgenbilder är en viktig del i planeringen av strålbehandling av patienter med hjärntumörer. Det höga antalet bilder och den nödvändiga precisionsnivån gör dock manuellsegmentering till en tidskrävande uppgift. Faltningsnätverk har därför föreslagits som ett verktyg förautomatiserad segmentering och visat lovande resultat. Datamängderna som används för att träna dessa djupinlärningsmodeller är ofta obalanserade och innehåller data som inte bidrar till modellensprestanda. Det finns därför potential att både skynda på träningen och förbättra nätverkets förmåga att segmentera tumörer genom att noggrant välja vilken data som används för träning. Denna uppsats implementerar importance sampling för att träna ett faltningsnätverk för patch-baserad segmentering av tredimensionella multimodala magnetröntgenbilder av hjärnan. Modellensträningstid och prestanda jämförs mot ett nätverk tränat med standardmetoden. Detta görs förtvå olika storlekar på patches. Egenskaperna hos de mest valda volymerna analyseras också. Importance sampling uppvisar en snabbare träningsprocess med avseende på antal epoker och resulterar också i modeller med högre prestanda. Analys av de oftast valda volymerna indikerar att under träning med stora patches förekommer små tumörer i en något högre utsträckning. Vidareundersökningar är dock nödvändiga för att bekräfta vilka aspekter som påverkar hur ofta en volym används.