Rozprawy doktorskie na temat „CLASSIFICATION OF BRAIN TUMOR”

A brain tumor is detected and classified by biopsy that is conducted after the brain surgery. Advancement in technology and machine learning techniques could help radiologists in the diagnosis of tumors without any invasive measures. We utilized a deep learning-based approach to detect and classify the tumor into Meningioma, Glioma, Pituitary tumors. We used registration and segmentation-based skull stripping mechanism to remove the skull from the MRI images and the grab cut method to verify whether the skull stripped MRI masks retained the features of the tumor for accurate classification. In this research, we proposed a transfer learning based approach in conjunction with discriminative learning rates to perform the classification of brain tumors. The data set used is a 3064 T MRI images dataset that contains T1 flair MRI images. We achieved a classification accuracy of 98.83%, 96.26%, and 95.18% for training, validation, and test sets and an F1 score of 0.96 on the T1 Flair MRI dataset.

The job of classifying or annotating brain tumors from MRI images can be time-consuming and difficult, even for radiologists. To increase the survival chances of a patient, medical practitioners desire a means for quick and accurate diagnosis. While datasets like CIFAR, ImageNet, and SVHN have tens of thousands, hundreds of thousands, or millions of samples, an MRI dataset may not have the same luxury of receiving accurate labels for each image containing a tumor. This work covers three models that classify brain tumors using a combination of convolutional neural networks and of the concept of capsule layers. Each network utilizes a hit-or-miss capsule layer to relate classes to capsule vectors in a one-to-one relationship. Additionally, this work proposes the use of deep active learning for picking the samples that can give the best model, PSP-HitNet, the most information when adding mini-batches of unlabeled data into the master, labeled training dataset. By using an uncertainty estimated querying strategy, PSP-HitNet approaches the best validation accuracy possible within the first 12-24% of added data from the unlabeled dataset, whereas random choosing takes until 30-50% of the unlabeled to reach the same performance.

Histopathology refers to the visual inspection of tissue under the microscope and it is the core part of diagnosis. The process of manual inspection of histopathology slides is very time-consuming for pathologists and error-prone. Furthermore, diagnosis can sometimes differ among specialists. In recent years, convolutional neural networks (CNNs) have demonstrated remarkable performances in the classification of digital histopathology images. However, due to their high resolution, whole-slide images are of immense size, often gigapixels, making it infeasible to train CNNs directly on them. For that, patch-level classification is used instead. In this study, a deep learning approach for patch-level classification of glioblastoma (i.e. brain cancer) tissue is proposed. Four different state-of-the-art models were evaluated (MobileNetV2, ResNet50, ResNet152V2, and VGG16), with MobileNetV2 being the best among them, achieving 80% test accuracy. The study also proposes a scratch-trained CNN architecture, inspired by the popular VGG16 model, which achieved 81% accuracy. Both models scored 87% test accuracy when trained with data augmentation. All models were trained and tested on randomly sampled patches from the Ivy GAP dataset, which consisted of 724 H&E images in total. Finally, as patch-level predictions cannot be used explicitly by pathologists, prediction results from two slides were presented in the form of whole-slide images. Post-processing was also performed on those two predicted WSIs in order to make use of the spatial correlations among the patches and increase the classification accuracy. The models were statistically compared using the Wilcoxon signed-rank test.

Brain and nervous system tumors were responsible for around 250,000 deaths in 2020 worldwide. Correctly identifying different tumors is very important, because treatment options largely depend on the diagnosis. This is an expert task, but recently machine learning, and especially deep learning models have shown huge potential in tumor classification problems, and can provide fast and reliable support for pathologists in the decision making process. This thesis investigates classification of two brain tumors, glioblastoma multiforme and lower grade glioma in high-resolution H&E-stained histology images using deep learning. The dataset is publicly available from TCGA, and 220 whole slide images were used in this study. Ground truth labels were only available on whole slide level, but due to their large size, they could not be processed by convolutional neural networks. Therefore, patches were extracted from the whole slide images in two sizes and fed into separate networks for training. Preprocessing steps ensured that irrelevant information about the background was excluded, and that the images were stain normalized. The patch-level predictions were then combined to slide level, and the classification performance was measured on a test set. Experiments were conducted about the usefulness of pre-trained CNN models and data augmentation techniques, and the best method was selected after statistical comparisons. Following the patch-level training, five slide aggregation approaches were studied, and compared to build a whole slide classifier model. Best performance was achieved when using small patches (336 x 336 pixels), pre-trained CNN model without frozen layers, and mirroring data augmentation. The majority voting slide aggregation method resulted in the best whole slide classifier with 91.7% test accuracy and 100% sensitivity. In many comparisons, however, statistical significance could not be shown because of the relatively small size of the test set.

Tumors have several important hallmarks; anomalous and heterogeneous behaviors of their vascular structures, and high angiogenesis and neovascularization. Tumor tissue presents high blood flow (F) and extraction ratio (E) of contrast molecules. Consequently there is growing interest in non invasive methods for characterizing changes in tumor vasculature. Toft's model has been extensively used in the past in order to calculate Ktrans maps which take into consideration both F and E. However, in this thesis we argue that for accurate tumor characterization we need a model able to compute both F and E in tissue plasma. This project has been developed as part of a larger project, working toward building a Clinical Decision Support System (CDSS): an interactive expert computer software, that helps doctors and other health professionals make decisions regarding patient treatment progress. Using the Gamma Capillary Transit Time (GCTT) pharmacokinetic model we calculate F and E separately in a more realistic framework; unlike other models it takes into account the heterogeneity of the tumor, which depends on parameter a-1. a-1 is the width of the distribution of the capillary transit times within a tissue voxel. In more detail, a-1 expresses the heterogeneity of tissue microcirculation and microvasculature. We studied 9 patients pathologically diagnosed with glioblastoma multiforme (GBM), a common malignant type of brain tumor. Several physiological parameters including the blood flow and extraction ratio distributions were calculated for each patient. Then we investigated if these parameters can characterize early the patients' responsiveness to current treatment; we assessed the classification potential based on the actual therapy outcome. To this end, we present a novel analysis framework which exploits the new parameter a-1 and organizes each voxel into four sub-region. Our results indicate that early characterization of response based on GCCT can be significantly improved by focusing on tumor voxels from a specific sub-region.

Malignant brain tumors, especially malignant gliomas, have a poor prognosis, a fact which has remained unchanged over the last decades despite the employment of multimodal therapeutic approaches. Malignant gliomas are among the most vascularized tumors known and the amount of vascularization has been correlated to their prognosis. Since tumor growth is dependent on concomitant vascularization, recent experimental studies have focused on the use of anti-angiogenic molecules as a novel strategy in brain tumor therapy. Angiogenesis inhibitors target at proliferating endothelial cells and suppress the formation of a sufficient vascular bed. Inhibitors such as TNP-470, suramin and angiostatin have shown their therapeutic potential in experimental studies. In a clinical setting, they could be applied for the treatment of multiple tumors or postsurgically as an adjuvant therapy to prevent recurrence. This article discusses presently available anti-angiogenic agents, emphasizing on substances already in clinical trials.
Maligne Hirntumoren, insbesondere die malignen Gliome, haben trotz multimodaler Therapieansätze eine unverändert schlechte Prognose. Diese Aggressivität korreliert mit der Tatsache, daß maligne Gliome zu den gefäßreichsten Tumoren zählen, die wir kennen. Die Quantifizierung der Gefäßdichte in diesen Tumoren erlaubte die Korrelation zur Überlebenszeit der Patienten. Da das Tumorwachstum von einer begleitenden Neovaskularisierung abhängt, wurden erste experimentelle Therapieansätze durchgeführt, um das Tumorwachstum durch Inhibierung der Neovaskularisierung zu verhindern. Inhibitoren der Angiogenese, z.B. TNP-470, Suramin und Angiostatin hemmen die Proliferation von Endothelzellen und die Ausbildung eines funktionsfähigen Gefäßbettes. Erste experimentelle Ansätze haben ihre tumorstatische Wirksamkeit in vivo bewiesen. Zur klinischen Behandlung wären diese Substanzen in Verbindung mit bestehenden Therapien einsetzbar, insbesondere für die Behandlung multipler Tumoren und zur postoperativen Therapie. Diese Übersichtsarbeit beschreibt die neuesten anti-angiogenen Therapiekonzepte besonders mit Hinblick auf Substanzen, die in ersten klinischen Studien eingesetzt werden.
Dieser Beitrag ist mit Zustimmung des Rechteinhabers aufgrund einer (DFG-geförderten) Allianz- bzw. Nationallizenz frei zugänglich.

Detecting and diagnosing brain tumour types quickly and accurately is essential to any effective treatment. The general brain tumour diagnosis procedure, biopsy, not only causes a great deal of pain to the patient but also raises operational difficulty to the clinician. In this thesis, a non-invasive brain tumour diagnosis system based on MR images is proposed. The first part is image preprocessing applied to original MR images from the hospital. Non-uniformed intensity scales of MR images are standardized relying on their statistic characteristics without requiring prior or post templates. It is followed by a non-brain region removal process using morphologic operations and a contrast enhancement between white matter and grey matter by means of histogram equalization. The second part is image segmentation applied to preprocessed MR images. A new image segmentation algorithm named IFCM is developed based on the traditional FCM algorithm. Neighbourhood attractions considered in IFCM enable this new algorithm insensitive to noise, while a neural network model is designed to determine optimized degrees of attractions. This extension can also estimate inhomogenities. Brain tissue intensities are acquired from segmentation. The final part of the system is brain tumour classification. It extracts hidden diagnosis information from brain tissue intensities using a fuzzy logic based GP algorithm. This novel method imports a fuzzy membership to implement a multi-class classification directly without converting it into several binary classification problems as with most other methods. Two fitness functions are defined to describe the features of medical data precisely. The superiority of image analysis methods in each part was demonstrated on synthetic images and real MR images. Classification rules of three types and two grades of brain tumours were discovered. The final diagnosis accuracy was very promising. The feasibility and capability of the non-invasive diagnosis system were testified comprehensively.

Multi-spectral images have the advantage of providing complementary information to resolve some ambiguities. But, the challenge is how to make use of the multi-spectral images effectively. In this thesis, our study focuses on the fusion of multi-spectral images by extracting the most useful features to obtain the best segmentation with the least cost in time. The Support Vector Machine (SVM) classification integrated with a selection of the features in a kernel space is proposed. The selection criterion is defined by the kernel class separability. Based on this SVM classification, a framework to follow up brain tumor evolution is proposed, which consists of the following steps: to learn the brain tumors and select the features from the first magnetic resonance imaging (MRI) examination of the patients; to automatically segment the tumor in new data using a multi-kernel SVM based classification; to refine the tumor contour by a region growing technique; and to possibly carry out an adaptive training. The proposed system was tested on 13 patients with 24 examinations, including 72 MRI sequences and 1728 images. Compared with the manual traces of the doctors as the ground truth, the average classification accuracy reaches 98.9%. The system utilizes several novel feature selection methods to test the integration of feature selection and SVM classifiers. Also compared with the traditional SVM, Fuzzy C-means, the neural network and an improved level set method, the segmentation results and quantitative data analysis demonstrate the effectiveness of our proposed system.

A lo largo de las últimas décadas, la disponibilidad cada vez mayor de grandes cantidades de información biomédica ha potenciado el desarrollo de herramientas que permiten extraer e inferir conocimiento. El aumento de tecnologías biomédicas que asisten a los expertos médicos en sus decisiones ha contribuido a la incorporación de un paradigma de medicina basada en la evidencia centrada en el paciente. Las contribuciones de esta Tesis se centran en el desarrollo de herramientas que asisten al médico en su proceso de toma de decisiones en el diagnóstico de tumores cerebrales (TC) mediante Espectros de Resonancia Magnética (ERM). En esta Tesis se contribuye con el desarrollo de clasificadores basados en Reconocimiento de Patrones (RP) entrenados con ERM de pacientes pediátricos y adultos para establecer el tipo y grado del tumor. Estos clasificadores especializados son capaces de aprovechar las diferencias bioquímicas existentes entre los TC infantiles y de adultos. Una de las principales contribuciones de esta Tesis consiste en el desarrollo de modelos de clasificación enfocados a discriminar los tres tipos de tumores cerebrales pediátricos más prevalentes. El cerebelo suele ser una localización habitual de estos tumores, resultando muy difícil distinguir el tipo mediante el uso de Imagen de Resonancia Magnética. Por lo tanto, obtener un alto acierto en la discriminación de astrocitomas pilocíticos, ependimomas y meduloblastomas mediante ERM resulta crucial para establecer una estrategia de cirugía, ya que cada tipo de tumor requiere de unas acciones diferentes si se quiere obtener un buen pronóstico del paciente. Asimismo, esta Tesis contribuye en la extracción de características para ERM mediante el estudio de la combinación de señales de ERM adquiridas en dos tiempos de eco: tiempo de eco corto y tiempo de eco largo; concluyendo que dicha combinación mejora la clasificación de tumores cerebrales pediátricos frente al hecho de usar únicamente los ERM de un
Vicente Robledo, J. (2012). Clinical Decision Support Systems for Brain Tumour Diagnosis: Classification and Evaluation Approaches [Tesis doctoral]. Editorial Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/17468
Palancia

Medical imagery is increasingly evolving towards higher resolution and throughput. The increasing volume of data and the usage of multiple and often novel imaging modalities necessitates the use of mathematical and computational techniques for quicker, more accurate and more robust analysis of medical imagery. The fields of computer vision and machine learning provide a rich set of techniques that are useful in medical image analysis, in tasks ranging from segmentation to classification and population analysis, notably by integrating the qualitative knowledge of experts in anatomy and the pathologies of various disorders and making it applicable to the analysis of medical imagery going forward. The object of the proposed research is exactly to explore various computer vision and machine learning methods with a view to the improved analysis of multiple modalities of brain and cardiac imagery, towards enabling the clinical goals of studying schizophrenia, brain tumors (meningiomas and gliomas in particular) and cardiovascular disorders. In the first project, a framework is proposed for the segmentation of tubular, branched anatomical structures. The framework uses the tubular surface model which yields computational advantages and further incorporates a novel automatic branch detection algorithm. It is successfully applied to the segmentation of neural fiber bundles and blood vessels. In the second project, a novel population analysis framework is built using the shape model proposed as part of the first project. This framework is applied to the analysis of neural fiber bundles towards the detection and understanding of schizophrenia. In the third and final project, the use of mass spectrometry imaging for the analysis of brain tumors is motivated on two fronts, towards the offline classification analysis of the data, as well as the end application of intraoperative detection of tumor boundaries. SVMs are applied for the classification of gliomas into one of four subtypes towards application in building appropriate treatment plans, and multiple statistical measures are studied with a view to feature extraction (or biomarker detection). The problem of intraoperative tumor boundary detection is formulated as a detection of local minima of the spatial map of tumor cell concentration which in turn is modeled as a function of the mass spectra, via regression techniques.

De nos jours, obtenir une segmentation efficace des tumeurs cérébrales de Glioblastome Multiforme (GBM) dans des images IRM multimodale le plus tôt possible, donne un diagnostic clinique, traitement et suivi précoce. La technique d'IRM est conçue spécifiquement pour fournir aux radiologues des outils puissants de visualisation pour analyser des images médicales, mais le challenge réside dans l'interprétation des images radiologiques avec les données cliniques et pathologiques et leurs causes dans les tumeurs GBM. C'est pourquoi la recherche quantitative en neuroimagerie nécessite souvent une segmentation anatomique du cerveau humain à partir d'images IRM afin d'aider la détection et la segmentation des tumeurs cérébrales. L'objectif de cette thèse est de proposer des méthodes automatiques de Deep learning pour la segmentation des tumeurs cérébrales à l'aide des images IRM.Tout d’abord, nous nous intéressons principalement à la segmentation des images IRM des patients atteints des tumeurs GBM en utilisant le Deep learning, en particulier, Deep Convolutional Neural Networks (DCNNs). Nous proposons deux approches End-to-End DCNNs pour la segmentation automatique des tumeurs cérébrales. La première approche est basée sur la technique pixel-wise et la deuxième approche est basée sur la technique patch-wise. Ensuite, nous prouvons que la deuxième approche est plus efficace en termes de performance de segmentation et de temps de calcul. Nous proposons aussi un nouvel algorithme d'optimisation pour optimiser les hyperparamètres adaptés à la première approche. Deuxièmement, pour améliorer les performances de segmentation des approches proposées, nous proposons de nouveaux pipelines de segmentation des images IRM des patients, où ces pipelines sont basés sur des features extraites de DCNNs et de deux étapes de training. Nous abordons aussi les problèmes liés aux données déséquilibrées en plus les faux positifs et les faux négatifs pour augmenter la sensibilité de segmentation du modèle vers les régions tumorales et la spécificité vers les régions saines. Finalement, les performances et le temps de segmentation des approches et des pipelines proposés sont rapportés avec les méthodes de l'état de l'art sur une base de données accessible au public, annotées par des radiologues et approuvées par des neuroradiologues
Nowadays, getting an efficient segmentation of Glioblastoma Multiforme (GBM) braintumors in multi-sequence MRI images as soon as possible, gives an early clinical diagnosis, treatment, and follow-up. The MRI technique is designed specifically to provide radiologists with powerful visualization tools to analyze medical images, but the challenge lies more in the information interpretation of radiological images with clinical and pathologies data and their causes in the GBM tumors. This is why quantitative research in neuroimaging often requires anatomical segmentation of the human brain from MRI images for the detection and segmentation of brain tumors. The objective of the thesis is to propose automatic Deep Learning methods for brain tumors segmentation using MRI images.First, we are mainly interested in the segmentation of patients’ MRI images with GBMbrain tumors using Deep Learning methods, in particular, Deep Convolutional NeuralNetworks (DCNN). We propose two end-to-end DCNN-based approaches for fully automaticbrain tumor segmentation. The first approach is based on the pixel-wise techniquewhile the second one is based on the patch-wise technique. Then, we prove that thelatter is more efficient in terms of segmentation performance and computational benefits. We also propose a new guided optimization algorithm to optimize the suitable hyperparameters for the first approach. Second, to enhance the segmentation performance of the proposed approaches, we propose new segmentation pipelines of patients’ MRI images, where these pipelines are based on deep learned features and two stages of training. We also address problems related to unbalanced data in addition to false positives and false negatives to increase the model segmentation sensitivity towards the tumor regions and specificity towards the healthy regions. Finally, the segmentation performance and the inference time of the proposed approaches and pipelines are reported along with state-of-the-art methods on a public dataset annotated by radiologists and approved by neuroradiologists

Malignant gliomas are the most common type of primary brain tumors in adults. Contrarily to brain metastases that have clear borders, malignant gliomas are poorly circumscribed lesions because they diffusely infiltrate the brain parenchyma. When possible, standard treatment of gliomas includes surgical resection, though surgeons unintentionally leave behind some of the tumor more than 50% of the time. Two factors are mainly responsible for this situation. First, most current neuronavigation systems are based on preoperative images; these systems become less accurate as the surgery progresses and the brain goes through important changes and deformations. Second, glioma boundaries are often visually and haptically difficult to determine. This is an unfortunate situation since maximum safe resection of these tumors correlates with longer survival times in patients presenting either a low-grade or high-grade glioma. By providing real-time images, intraoperative imaging techniques aid neurosurgeons achieve more complete resections while also helping to prevent damage to normal brain. In this thesis, I have investigated the use of intraoperative ultrasound to guide glioma surgery. To achieve this, I have used the prototype neuronavigation system developed by our research group: the IBIS NeuroNav system.The aim of the first paper was to evaluate the precision and accuracy of IBIS NeuroNav. Four aspects of the system were characterized: 1) the ultrasound probe calibration, 2) the temporal calibration, 3) the patient-to-image registration and 4) the mean intial MRI-ultrasound misalignment. IBIS NeuroNav was found to have an accuracy similar to other comparable systems in the literature.The goal of the second paper was to present a new technique for the rigid registration of the preoperative MRI to the pre-resection intraoperative ultrasound. Initially these images generally have a slight misalignment. However, surgeons find ultrasound images easier to interpret when they are properly aligned with MRI. The results of our investigation showed that the proposed registration technique robustly improved the MRI–ultrasound alignment when compared with the initial alignment.The objective of the third paper was to test rigid and non-rigid registration techniques to better align the pre- and post-resection ultrasound images to facilitate interpretation of the latter. A simple correlation coefficient based nonlinear registration proved to significantly improve the alignment between the pre- and post-resection ultrasound images. One of the biggest challenges of many technical scientists in the field of medical imaging is to find clinical images on which to validate new image processing algorithms. To address this issue, the fourth paper presents an online database in which we share our acquired preoperative MRI and intraoperative ultrasound images with the medical imaging community. We hope that this work will make it easier for neurosurgeons to use intraoperative ultrasound to guide glioma surgery and will eventually lead to improved surgical accuracy and prolonged patient survival.
Les gliomes malins sont les tumeurs primaires les plus répandues chez l'adulte. Contrairement aux métastases cérébrales qui ont des contours bien définis, les gliomes malins infiltre le cerveau environnant et ont souvent des contours plus indistincts. En présence d'un gliome, en général la résection chirurgicale est l'approche privilégiée. Par ailleurs, dans un cas sur deux les neurochirurgiens laissent involontairement une partie de la tumeur. Deux facteurs principaux sont responsables de cet état de fait. Premièrement, la plupart des systèmes de neuronavigation actuels sont basés sur des images préopératoires. Parce que le cerveau subit d'importants changements lors de la chirurgie, ces images perdent en précision au fil de l'opération. En second lieu, les limites d'un gliome sont souvent difficiles à déterminer de façon précise, à la fois avec le sens de la vue et du toucher. Cette situation est regrettable puisqu'une résection à la fois maximale et sécuritaire de ces tumeurs est corrélée avec une survie prolongée des patients présentant un gliome de bas ou de haut grade. L'imagerie peropératoire permet d'obtenir des images en temps réel, aidant ainsi le chirurgien à faire une résection plus complète, tout en protégeant les tissus sains. Dans cette thèse j'ai étudié l'usage de l'ultrason peropératoire afin de guider la résection d'un gliome. À cette fin, j'ai utilisé le prototype de système de neuronavigation développé dans notre groupe de recherche : le système IBIS NeuroNav. Le but du premier article était d'évaluer la précision d'IBIS NeuroNav. Quatre composants du système ont été considérés : 1) le calibrage de la sonde ultrason 2) le calibrage temporel 3) le recalage patient-image et 4) le recalage IRM-ultrason. Nous avons constaté qu'IBIS NeuroNav avait une précision similaire aux autres systèmes comparables présentés dans la littérature. Le but du deuxième article était de présenter une nouvelle technique de recalage rigide entre l'IRM préopératoire et l'ultrason pré-résection intraopératoire. Au départ, ces images sont généralement légèrement désalignées. Or, les chirurgiens trouvent l'interprétation des images ultrasons plus facile lorsqu'elles sont correctement alignées avec l'IRM. Nos résultats montrent que la nouvelle technique proposée améliore de façon significative l'alignement IRM-ultrason.Le but du troisième article était de tester des méthodes de recalage rigide et non-rigide pour améliorer l'alignement des images ultrasons pré- et post-opératoires, afin de faciliter l'interprétation des seconds. Nous avons trouvé qu'une méthode de recalage utilisant un simple coefficient de corrélation améliorait significativement cet alignement. Un des nombreux défis des scientifiques techniques du domaine de l'imagerie médicale est de trouver des images cliniques sur lesquelles valider leurs nouveaux algorithmes. L'objectif du quatrième papier était précisément de pallier à cette difficulté en partageant avec la communauté scientifique les images acquises pour les papiers précédents. Nous sommes confiants que les résultats présentés dans cette thèse faciliteront l'utilisation par les neurochirurgiens des ultrasons peropératoires. Une survie prolongée de trop nombreux patients en dépend!

Conselho Nacional de Desenvolvimento CientÃfico e TecnolÃgico
CoordenaÃÃo de AperfeiÃoamento de Pessoal de NÃvel Superior
The disabling effects of central nervous system (CNS) tumors are out of proportion to their low incidence. Theyâre second only to stroke as neurologic mortality causes. Brain metastases are the commonest intracranial tumors in adults, almost 10 times more frequent than primary brain tumors. Their diagnosis and treatment have met significant advances, although much more research about drug resistance and new treatment modalities are needed. New and even better brain tumor animal models will help to evaluate novel drug regimens and adjuvant therapies for CNS neoplasms. In the present work, the author presents a simple and easily reproducible brain tumor model utilizing the tumor cell line W256 transplanted to Wistar rats. They tested a drug widely used for palliative treatment of tumoral brain edema (dexamethasone), for survival impact. They also have tested the effects of a drug newly proposed as multidrug resistance reversal agent (cyclosporin â CS). Wistar albino rats had stereotaxic intracranial tumor inoculation after the surgical installation of a permanent canulla on the area of interest (right subfrontal caudate). The brain tumor model, as a model of metastatic brain disease, was successful, with reproducible tumor growth (95%), low incidence of extracranial tumor implantation (21% local, no distant metastasis) and few evidence of surgical site infection (21%). The median survival of the animals was 12.5 days (control), 13 days (CS vehicle treated), 11 days (CS treated), 9.5 and 9 days (dexamethasone 0.3 and 3.0 mg/kg/day). These differences were not significant, although the survival rates on the 12th day post-inoculation have showed a significant survival decrease for the case of dexamethasone 3,0 mg/kg/day (p < 0.05), but not for CS treatment (Fischerâs Exact Test). The estimated tumor volume was 17.08 Â 6.7 mm3 (control) and 12.61 Â 3.6 mm3 (CS treatment, not significant, Studentâs t-test). The tumor volume in the 9th day post-inoculation was estimated in 67,25 Â 19,8 mm3. The doubling time was 24.25 h. This model behaved as an undifferentiated tumor, with local invasiveness features compared with that of primary brain tumors. It fits well, in this way, for the study of tumor cell migration on CNS parenchyma. Phenomena like neuronal degeneration, neuron cell edema and death, and gliosis, as well as perivascular cell infiltrates, were seen frequently. One could find, also, neoangiogenesis, satellite tumor growth, and tumor cell migration in normal brain parenchyma. Besides heavy parenchymatous infiltration, it was also disclosed markedly tumor cell migration along white matter tracts, such as callosal fibers and infiltration in the Virchow-Robins perivascular space. The model presents as a dual brain tumor and leptomeningeal carcinomatosis model. It could be used for the study and treatment test in the scenario of these two pathologies. The intracerebral tumor growth induced peripheral blood neutrophil count elevation (ANOVA, p < 0.01), higher chance for neutrophilia (Fischerâs Exact Test, p < 0.01), higher chance for lymphopenia (Fischerâs Exact Test, p < 0.01) and brain weight increase (Studentâs t-test, p < 0.001) comparing to control. There was no significant change in any of the other hematologic, biochemical and biological parameters tested. CS treatment did not alter any of the tests, as compared to non-treated brain tumor animals. The only exception was the mean animal weight on the first week post-inoculation (ANOVA, p < 0.05). CS, in this way, was responsible for an early cachexia in the brain tumor inoculated animals. CS treatment of brain tumor animals did show non-significant effects indicating a volume (26%) and weight tumor decrease, and tumor infiltrating neutrophil increase (odds ratio - OR = 5.6). This indicates the necessity to further study morphologically and functionally the local inflammation in brain tumor inoculated animals, as well the effects of CS administration. In conclusion, the W256 intracerebral tumor model is simple, easily performed, reproducible and of great potential utility. In this model, tumor inoculation can lead to hematologic and biologic modifications in the experimental animals. CS could apparently lead to early tumor caquexia in this tumor model. However, CS treatment did not modify the survival chance of the brain tumor animals, in sharp contrast to dexamethasone 3.0mg/kg/day, a much-used drug in the treatment of brain tumors, which decreased the animal survival.
Os importantes efeitos incapacitantes dos tumores do sistema nervoso central (SNC) sÃo desproporcionais a sua baixa incidÃncia. Mesmo assim, entre as doenÃas neurolÃgicas, ficam atrÃs apenas dos acidentes vasculares do SNC como causa de morte. MetÃstases cerebrais constituem os tumores intracranianos mais comuns do adulto, ocorrendo atà 10 vezes mais freqÃentemente que tumores primÃrios. AvanÃos significativos ocorreram em seu diagnÃstico e tratamento, embora mais pesquisa sobre os fenÃmenos que diminuem o efeito de drogas em metÃstases cerebrais e tratamentos eficazes para estas patologias sejam cada vez mais necessÃrios. O desenvolvimento de melhores modelos animais de tumores do SNC serà necessÃrio para a avaliaÃÃo in vivo de novas formas de quimioterapia (QT) e terapia adjuvante para tumores cerebrais. No presente trabalho, o autor objetivou desenvolver um modelo de tumor cerebral simples e de fÃcil reproduÃÃo utilizando a linhagem W256, alÃm de testar o efeito na sobrevida animal de uma droga largamente usada para o tratamento de efeitos secundÃrios a edema cerebral (dexametasona). O autor tambÃm testou uma droga envolvida numa nova proposta de reversÃo de multirresistÃncia a drogas anti-neoplÃsicas em tumores cerebrais (ciclosporina â CS). Ratos albinos (Wistar) tiveram o tumor inoculado atravÃs de estereotaxia, apÃs a instalaÃÃo cirÃrgica de uma cÃnula no ponto escolhido (caudato subfrontal direito). O modelo de tumor implantado no cÃrebro de ratos, simulando uma metÃstase cerebral, mostrou-se bem sucedido e reprodutÃvel (95% de crescimento tumoral), com baixa incidÃncia de disseminaÃÃo tumoral extracraniana local (21%), baixa evidÃncia de infecÃÃo local (21%), ausÃncia de metÃstases à distÃncia e ausÃncia de sinais de infecÃÃo sistÃmica. Os animais sobreviveram uma mediana de 12,5 dias (grupo controle), 13 dias (tratados com veÃculo da CS), 11 dias (tratados com CS), 9,5 e 9 dias (dexametasona 0,3 e 3,0 mg/kg/dia, respectivamente). As diferenÃas entre estas medianas nÃo foram significantes (teste de Kruskal-Wallis), embora as diferenÃas entre as taxas de sobrevida no 12o dia apÃs a inoculaÃÃo tenham mostrado reduÃÃo significante no grupo que recebeu dexametasona 3,0 mg/kg/dia (p < 0,05), mas nÃo no grupo tratado com CS (teste de Fischer). O volume tumoral estimado (VTE) no sÃtimo dia pÃs-inoculaÃÃo (7DPI) foi de 17,08  6,7 mm3 no controle e 12,61 3,6 mm3 apÃs tratamento com CS, sem diferenÃa significante (teste t-Student). O VTE no 9DPI de animais do grupo Tumor foi de 67,25  19,8 mm3. O tempo de duplicaÃÃo foi de 24,25 h. O modelo comportou-se como um tumor de caracterÃsticas indiferenciadas, apresentando invasividade local comparada à de tumores primÃrios do SNC, prestando-se ao estudo da migraÃÃo de cÃlulas tumorais no SNC. Observaram-se fenÃmenos como degeneraÃÃo neuronal hidrÃpica, edema celular neuronal, sinais de morte celular neuronal e gliose, alÃm da presenÃa de infiltrados celulares tumorais e inflamatÃrios perivasculares. Observaram-se, tambÃm, neoformaÃÃo vascular, formaÃÃo de nÃdulos tumorais satÃlites ao tumor principal e migraÃÃo celular tumoral no parÃnquima cerebral normal. Observou-se, alÃm da infiltraÃÃo parenquimatosa, marcante migraÃÃo celular tumoral ao longo de tratos de substÃncia branca (corpo caloso) e ao longo dos espaÃos perivasculares de Virchow-Robins. O modelo apresenta-se como um misto de tumor cerebral intraparenquimatoso e carcinomatose leptomenÃngea, podendo ser utilizado para estudar o comportamento e testar formas de tratamento para ambas as patologias. O crescimento tumoral intracerebral induziu aumento do nÃmero de neutrÃfilos no sangue perifÃrico (ANOVA, p < 0,01), maior chance de apresentar neutrofilia (teste de Fischer, p < 0,01), maior chance de apresentar linfopenia (teste de Fischer, p < 0,01) e aumento do peso dos cÃrebros dos animais experimentais (teste t-Student, p < 0,001) em relaÃÃo ao controle. Nenhum dos outros valores hematolÃgicos, bioquÃmicos e biolÃgicos foi alterado de maneira significante. O tratamento de animais inoculados com tumor com a CS, nÃo alterou nenhuma das medidas hematolÃgicas, bioquÃmicas ou biolÃgicas em relaÃÃo aos animais inoculados com tumor e nÃo tratados, exceto o peso dos animais na primeira semana apÃs inoculaÃÃo tumoral (ANOVA, p < 0,05). A CS, dessa forma, induziu significantemente uma caquexia precoce nos animais inoculados com tumor cerebral. O tratamento com CS de animais inoculados com tumor mostrou tendÃncias nÃo significantes a diminuir volume (26%) e massa (7%) tumorais e aumentar nÃmero de neutrÃfilos infiltrantes de tumor (razÃo de chance - RC = 5,6) e necrose tumoral, indicando a necessidade de posteriores estudos para caracterizar morfolÃgica e funcionalmente a resposta inflamatÃria local em animais inoculados com tumor e a influÃncia da CS neste processo, alÃm do efeito da CS na angiogÃnese tumoral. Concluindo, o modelo de W256 intracerebral mostrou-se simples, de fÃcil execuÃÃo, reprodutÃvel e Ãtil. Neste modelo, a inoculaÃÃo tumoral induz modificaÃÃes hematolÃgicas e biolÃgicas nos animais. A CS pareceu exarcebar a caquexia tumoral neste modelo. A CS, todavia, nÃo alterou a chance de sobrevida de animais inoculados com tumor cerebral, ao contrÃrio da dexametasona 3,0 mg/kg/dia, que reduziu esta chance. A CS, assim, parece ser mais segura neste modelo tumoral que uma droga largamente utilizada para tratamento de pacientes com metÃstase cerebral.

To prevent further brain tumour growth, malignant tissue should be removed as completely as possible in neurosurgical operations. Therefore, differentiation between tumour and brain tissue as well as detecting functional areas is very important. Hyperspectral imaging (HSI) can be used to get spatial information about brain tissue types and characteristics in a quasi-continuous reflection spectrum. In this paper, workflow and some aspects of an adapted hardware system for intraoperative hyperspectral data acquisition in neurosurgery are discussed. By comparing an intraoperative with a laboratory setup, the influences of the surgical microscope are made visible through the differences in illumination and a pixel- and wavelength-specific signal-to-noise ratio (SNR) calculation. Due to the significant differences in shape and wavelength-dependent intensity of light sources, it can be shown which kind of illumination is most suitable for the setups. Spectra between 550 and 1,000 nm are characterized of at least 40 dB SNR in laboratory and 25 dB in intraoperative setup in an area of the image relevant for evaluation. A first validation of the intraoperative hyperspectral imaging hardware setup shows that all system parts and intraoperatively recorded data can be evaluated. Exemplarily, a classification map was generated that allows visualization of measured properties of raw data. The results reveal that it is possible and beneficial to use HSI for wavelength-related intraoperative data acquisition in neurosurgery. There are still technical facts to optimize for raw data detection prior to adapting image processing algorithms to specify tissue quality and function.:Abstract Introduction Materials and methods (Clinical workflow and setup for hyperspectral imaging process, Characteristics of the lighting, Characteristics of the hyperspectral imaging camera, Spectral data acquisition and raw data pre-processing in neurosurgery, Spectral data evaluation) Results (Spectral characteristics of the lighting, SNR of the HSI camera, Data acquisition and raw data preprocessing during neurosurgical operation, Spectral data evaluation) Discussion Conclusions

A Thesis submitted to The University of Arizona College of Medicine - Phoenix in partial fulfillment of the requirements for the Degree of Doctor of Medicine.
Background: Great strides have been made in childhood cancer treatment efficacy over the past two decades leading to improved survival rates, and now attention is being directed toward identifying and understanding complications that affect many of these patients as they reach adulthood. Obesity is a well‐recognized late effect that has many potential long‐term consequences some of which include cardiovascular disease, type II diabetes mellitus, dyslipidemia and even death. Materials/Methods: We conducted a retrospective chart review to determine the prevalence of obesity among survivors of pediatric brain tumors 5 years after the completion of therapy and compare this to the general pediatric population of the same age. We also sought to identify potential risk factors for the development of obesity among survivors of childhood brain tumors. Obesity was defined as a body mass index (BMI) greater than the 95th percentile for age and gender as defined by the most recent Center for Disease Control growth curves. Results: We identified 96 patients who met our inclusion criteria, however only 43 had follow‐up data at 5 years after the completion of therapy to be included in final analysis. Of 43 patients, 5 (11.63%) were obese 5 years after completion of therapy. The CDC sites general population obesity rates in three age groups: 2‐5 years (8.4% obesity rate), 6‐ 11 years (18% obesity rate), 12‐19 years (21% obesity rate). Using CDC guidelines, we found no significant difference between the obesity rate among the brain tumor survivor population for each age group and the general population, p‐values of 0.865, 0.865, and 0.249 respectively. Conclusion: Our small sample size was likely not adequate to find a significant difference between the two groups or identify risk factors associated with the development of obesity. Larger studies are needed to further examine the risk of obesity among pediatric brain tumor survivors and to identify risk factors associated with this late effect.

Abstract : Malignant brain tumors are the second leading cause of cancer related deaths in children under 20. There are nearly 700,000 people in the U.S. living with a brain tumor and 17,000 people are likely to loose their lives due to primary malignant and central nervous system brain tumor every year. To identify whether a patient is diagnosed with brain tumor in a non-invasive way, an MRI scan of the brain is acquired followed by a manual examination of the scan by an expert who looks for lesions (i.e. cluster of cells which deviate from healthy tissue). For treatment purposes, the tumor and its sub-regions are outlined in a procedure known as brain tumor segmentation . Although brain tumor segmentation is primarily done manually, it is very time consuming and the segmentation is subject to variations both between observers and within the same observer. To address these issues, a number of automatic and semi-automatic methods have been proposed over the years to help physicians in the decision making process. Methods based on machine learning have been subjects of great interest in brain tumor segmentation. With the advent of deep learning methods and their success in many computer vision applications such as image classification, these methods have also started to gain popularity in medical image analysis. In this thesis, we explore different machine learning and deep learning methods applied to brain tumor segmentation.
Résumé: Les tumeurs malignes au cerveau sont la deuxième cause principale de décès chez les enfants de moins de 20 ans. Il y a près de 700 000 personnes aux États-Unis vivant avec une tumeur au cerveau, et 17 000 personnes sont chaque année à risque de perdre leur vie suite à une tumeur maligne primaire dans le système nerveu central. Pour identifier de façon non-invasive si un patient est atteint d'une tumeur au cerveau, une image IRM du cerveau est acquise et analysée à la main par un expert pour trouver des lésions (c.-à-d. un groupement de cellules qui diffère du tissu sain). Une tumeur et ses régions doivent être détectées à l'aide d'une segmentation pour aider son traitement. La segmentation de tumeur cérébrale et principalement faite à la main, c'est une procédure qui demande beaucoup de temps et les variations intra et inter expert pour un même cas varient beaucoup. Pour répondre à ces problèmes, il existe beaucoup de méthodes automatique et semi-automatique qui ont été proposés ces dernières années pour aider les praticiens à prendre des décisions. Les méthodes basées sur l'apprentissage automatique ont suscité un fort intérêt dans le domaine de la segmentation des tumeurs cérébrales. L'avènement des méthodes de Deep Learning et leurs succès dans maintes applications tels que la classification d'images a contribué à mettre de l'avant le Deep Learning dans l'analyse d'images médicales. Dans cette thèse, nous explorons diverses méthodes d'apprentissage automatique et de Deep Learning appliquées à la segmentation des tumeurs cérébrales.

Image guidance and visualization play an important role in modern surgery to help surgeons perform their surgical procedures. Here, the focus is on neurosurgery applications, in particular brain tumor surgery where a craniotomy (opening of the skull) is performed to access directly the brain region to be treated. In this type of surgery, once the skull is opened the brain can change its shape, and this deformation is known as brain shift. Moreover, the boundaries of many types of tumors are difficult to identify by the naked eye from healthy tissue. The main goal of this work was to study and develop image guidance and visualization methods for tumor surgery in order to overcome the problems faced in this type of surgery. Due to brain shift the magnetic resonance dataset acquired before the operation (preoperatively) no longer corresponds to the anatomy of the patient during the operation (intraoperatively). For this reason, in this work methods were studied and developed to compensate for this deformation. To guide the deformation methods, information of the superficial vessel centerlines of the brain was used. A method for accurate (approximately 1 mm) reconstruction of the vessel centerlines using a multiview camera system was developed. It uses geometrical constraints, relaxation labeling, thin plate spline filtering and finally mean shift to find the correct correspondences between the camera images. A complete non-rigid deformation pipeline was initially proposed and evaluated with an animal model. From these experiments it was observed that although the traditional non-rigid registration methods (in our case coherent point drift) were able to produce satisfactory vessel correspondences between preoperative and intraoperative vessels, in some specific areas the results were suboptimal. For this reason a new method was proposed that combined the coherent point drift and thin plate spline semilandmarks. This combination resulted in an accurate (below 1 mm) non-rigid registration method, evaluated with simulated data where artificial deformations were performed. Besides the non-rigid registration methods, a new rigid registration method to obtain the rigid transformation between the magnetic resonance dataset and the neuronavigation coordinate systems was also developed. Once the rigid transformation and the vessel correspondences are known, the thin plate spline can be used to perform the brain shift deformation. To do so, we have used two approaches: a direct and an indirect. With the direct approach, an image is created that represents the deformed data, and with the indirect approach, a new volume is first constructed and only after that can the deformed image be created. A comparison of these two approaches, implemented for the graphics processing units, in terms of performance and image quality, was performed. The indirect method was superior in terms of performance if the sampling along the ray is high, in comparison to the voxel grid, while the direct was superior otherwise. The image quality analysis seemed to indicate that the direct method is superior. Furthermore, visualization studies were performed to understand how different rendering methods and parameters influence the perception of the spatial position of enclosed objects (typical situation of a tumor enclosed in the brain). To test these methods a new single-monitor-mirror stereoscopic display was constructed. Using this display, stereo images simulating a tumor inside the brain were presented to the users with two rendering methods (illustrative rendering and simple alpha blending) and different levels of opacity. For the simple alpha blending method an optimal opacity level was found, while for the illustrative rendering method all the opacity levels used seemed to perform similarly. In conclusion, this work developed and evaluated 3D reconstruction, registration (rigid and non-rigid) and deformation methods with the purpose of minimizing the brain shift problem. Stereoscopic perception of the spatial position of enclosed objects was also studied using different rendering methods and parameter values.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2003.
Includes bibliographical references (p. 180-189).
A framework is proposed for the segmentation of brain tumors from MRI. Instead of training on pathology, the proposed method trains exclusively on healthy tissue. The algorithm attempts to recognize deviations from normalcy in order to compute a fitness map over the image associated with the presence of pathology. The resulting fitness map may then be used by conventional image segmentation techniques for honing in on boundary delineation. Such an approach is applicable to structures that are too irregular, in both shape and texture, to permit construction of comprehensive training sets. We develop the method of diagonalized nearest neighbor pattern recognition, and we use it to demonstrate that recognizing deviations from normalcy requires a rich understanding of context. Therefore, we propose a framework for a Contextual Dependency Network (CDN) that incorporates context at multiple levels: voxel intensities, neighborhood coherence, intra-structure properties, inter-structure relationships, and user input. Information flows bi-directionally between the layers via multi-level Markov random fields or iterated Bayesian classification. A simple instantiation of the framework has been implemented to perform preliminary experiments on synthetic and MRI data.
by David Thomas Gering.
Ph.D.

The DNA methylation is one of the main epigenetic inheritance form, which contributes in the regulation of gene expression. Abnormalities in DNA methylation processes can provide information about many pathophysiological conditions, including tumorigenesis. DNA hypomethylation was the initial epigenetic abnormality recognized in human tumors. Glioblastoma (GB) is the most common and the most aggressive primary brain tumor in adults and therefore is considered one of the major issue in modern medicine.

Children who have received radiation therapy for the treatment of brain tumors have been shown to experience neurocognitive deficits which appear to increase over time. The purpose of this study was to examine the memory functioning of 22 children irradiated for brain tumor and 22 healthy children of the same age who had not received irradiation. Subjects were administered a brief form of the WISC-III, to obtain an IQ, and the Wide Range Assessment of Memory and Learning (WRAML), to evaluate visual and verbal memory. Results indicated that, although there were no significant differences between the IQ scores of healthy children and children who had been irradiated, children who have received radiation therapy for brain tumor evidence memory deficits which effect visual and verbal memory abilities. Among the children who had been irradiated, as time since treatment increased, visual memory and overall memory functioning appeared to decline. Findings also suggested that children who received total tumor resection may evidence greater memory deficits than those who received only a partial resection. Visual memory was more closely related to IQ in the children irradiated for brain tumor than in the healthy children. The overall importance of research with this population lies in refining the understanding of memory deficits and strengths in order to formulate more effective remediation compensation, strategies.

Thesis advisor: Thomas N. Seyfried
Mitochondria are the key regulators of the bioenergetic state of the cell. Damage to mitochondrial protein, DNA, or membrane lipids can result as the cause or affect of disease pathology. Regardless, this damage can impair mitochondrial function resulting in a decreased ability to produce ATP to support cellular viability. This thesis research examined the mitochondrial lipidome by shotgun lipidomics in different populations of C57BL/6J (B6) brain mitochondria (non-synaptic and synaptic) and correlated lipid changes to differences in electron transport chain (ETC) activities. Furthermore, a comparison was made for non-synaptic mitochondria between the B6 and the VM mouse strain. The VM strain has a 1.5% incidence of spontaneous brain tumors, which is 210 fold greater than the B6 strain. I determined that differences in the brain mitochondrial lipidome existed in the VM strain compared to the B6 strain, likely corresponding to an increased rate of spontaneous brain tumor formation. Analysis of the mitochondrial genome in the CT-2A, EPEN, VM-NM1, and VM-M3 brain tumors compared to their syngeneic controls mouse strains, C57BL/6J (B6) and VM mice, was examined to determine if mutations existed in experimental brain cancer models. No pathogenic mtDNA mutations were discovered that would likely cause a decrease in the mitochondrial functionality. A novel hypothesis was devised to examine the tumor mitochondrial lipidome to determine if quantitative or molecular species differences existed that could potentially alter the functionality of the ETC. Brain tumor mitochondria were examined from tumors grown in vivo as well as in vitro. Numerous lipid differences were found in the mitochondria of brain tumors, of which the most interesting involved the unique molecular speciation of cardiolipin. ETC activities were significantly decreased in the primary ETC complexes which contribute protons to the gradient as well as the linked complexes of brain tumor mitochondria compared to controls. Taken together, it is likely that differences in the mitochondrial lipidome of brain tumors results in severe impairment of the mitochondria’s ability to produce ATP through the ETC. This research has provided a new understanding of the role of mitochondrial lipids in brain as well as brain cancer and offers an alternative explanation for metabolic dysfunction in cancer
Thesis (PhD) — Boston College, 2008
Submitted to: Boston College. Graduate School of Arts and Sciences
Discipline: Biology

Quantitative cancer imaging is an emerging field that develops computational techniques to acquire a deep understanding of cancer characteristics for cancer diagnosis and clinical decision making. The recent emergence of growing clinical imaging data provides a wealth of opportunity to systematically explore quantitative information to advance cancer diagnosis. Crucial questions arise as to how we can develop specific computational models that are capable of mining meaningful knowledge from a vast quantity of imaging data and how to transform such findings into improved personalized health care? This dissertation presents a set of computational models in the context of malignant brain tumors— Giloblastoma Multiforme (GBM), which is notoriously aggressive with a poor survival rate. In particular, this dissertation developed quantitative feature extraction approaches for tumor diagnosis from magnetic resonance imaging (MRI), including a multi-scale local computational feature and a novel regional habitat quantification analysis of tumors. In addition, we proposed a histogram-based representation to investigate biological features to characterize ecological dynamics, which is of great clinical interest in evaluating tumor cellular distributions. Furthermore, in regards to clinical systems, generic machine learning techniques are typically incapable of generalizing well to specific diagnostic problems. Therefore, quantitative analysis from a data-driven perspective is becoming critical. In this dissertation, we propose two specific data-driven models to tackle different types of clinical MRI data. First, we inspected cancer systems from a time-domain perspective. We propose a quantitative histogram-based approach that builds a prediction model, measuring the differences from pre- and post-treatment diagnostic MRI data. Second, we investigated the problem of mining knowledge from a skewed distribution—data samples of each survival group are unequally distributed. We proposed an algorithmic framework to effectively predict survival groups by jointly considering imbalanced distributions and classifier design. Our approach achieved an accuracy of 95.24%, suggesting it captures class-specific information in a challenging clinical setting.

MRI is favorable for brain imaging due to its excellent soft tissue contrast and absence of harmful ionizing radiation. Many have proposed supervised multimodal neural networks for automatic brain tumor segmentation and showed promising results. However, they rely on large amounts of labeled data to generalize well. The trained network is also highly specific to the task and input. Missing inputs will most likely have a detrimental effect on the network’s predictions, if it works at all. The aim of this thesis work is to implement a deep neural network that learns the general representation of multimodal MRI images in an unsupervised manner and is insensitive to missing modalities. With the latent representation, labeled data are then used for brain tumor segmentation. A variational autoencoder and an unified representation network are used for repre- sentation learning. Fine-tuning or joint training was used for segmentation task. The performances of the algorithms at the reconstruction task was evaluated using the mean- squared error and at the segmentation task using the Dice coefficient. Both networks demonstrated the possibility in learning brain MR representations, but the unified representation network was more successful at the segmentation task.

Thesis (M.S.) -- Worcester Polytechnic Institute.
Keywords: Visualization; Numerical analysis; Computational biology; Scientific computation; High-performance computing. Includes bibliographical references (p.19).

Thesis: Ph. D., Harvard-MIT Program in Health Sciences and Technology, 2014.
Cataloged from PDF version of thesis.
Includes bibliographical references (pages 115-127).
Brain tumor associated edema, a common feature of malignant brain neoplasms, is a significant cause of morbidity from brain tumor. Systemic administration of corticosteroids, the standard of care, is highly effective but can introduce serious systemic complications. Agents that inhibit the vascular endothelial growth factor (VEGF) pathway, such as cediranib, are promising alternatives, but are also associated with systemic toxicity as VEGF is essential for normal physiological functions. A miniature drug delivery device was developed for local drug delivery in rodents. It comprises of a drug reservoir and a cap with orifice(s) through which drug is released. Drug release kinetics is dependent on the payload, the drug solubility, and the surface area for diffusion. Sustained releases of dexamethasone (DXM), dexamethasone sodium phosphate (DSP), and solid dispersion of cediranib (AZD/PVP) were achieved. Employing the solid dispersion technique to increase the solubility of cediranib was necessary to enhance its release. Therapeutic efficacy and systemic toxicity of local drug administration via our devices were examined in an intracranial 9L gliosarcoma rat model. Local delivery of DSP was effective in reducing edema but led to DXM induced weight loss at high doses in a pilot study. DXM, which is much less water-soluble than DSP, was used subsequently to reduce the dose delivered. The use of DXM enabled long-term, sustained zero-order release and a higher payload than DSP. Local deliveries of DXM and AZD/PVP were demonstrated to be as effective as systemic dosing in alleviating edema. Edema reduction was associated with survival benefit, despite continuous tumor progression. Animals treated with locally delivered DXM did not suffer from body weight loss and corticosterone suppression, which are adverse effects induced by systemic DXM. Local drug administration using our device is superior to traditional systemic administration as it minimizes systemic toxicity and allows increased drug concentration in the tumor by circumventing the blood brain barrier. A much lower dose can therefore be utilized to achieve similar efficacy. Our drug delivery system can be used with other therapeutic agents targeting brain tumor to achieve therapeutic efficacy without systemic toxicity.
by Qunya Ong.
Ph. D.

Background: Glioblastomas are highly aggressive and malignant brain tumors which are difficult to resect totally. The surgical extent of resection constitutes a key role due to its direct influence on the patient’s survival time. To determine the resection extent, the tumor volume on pre-operative and post-operative magnetic resonance (MR) images should be calculated and compared. Materials and Methods: An active contour segmentation method was implemented to segment glioblastoma brain tumors on pre-operative T1-contrast enhanced MR images in axial, coronal and sagittal planes by self-developed software. The volume was rendered from the tumor  contours using Delaunay triangulation. Besides the segmentation and volume rendering, a graphical user interface was developed to facilitate the rendering, visualization and volume calculation of the brain tumor. The software was implemented in MATLAB (version 7.2). Two MR image data sets from glioblastoma patients were used and the repeatability and reproducibility of volume calculation was tested. Dimensions of the calculated tumor volume were then compared to the dimensions obtained in Amira® software. Results: The tumor volumes for data set 1 and data set 2 were 62.7 and 39.0 cm3, respectively. When tumor was segmented by different users (n=4), the volumes were 62.5 ± 0.3 and 42.6 ± 3.5 cm3. Segmentation errors were seen during the segmentation of data set 2. Mainly under- and over-segmentation due to hypointense MR signals caused by cerebrospinal fluid, or hyperintense MR signals caused by skull bone and weak tumor boundaries led to wrong segmentation results. Conclusion: Segmentation using active contours method is able to detect the brain tumor boundaries. The volume rendering and visualization allows the user to explore the tumor tissue and its surrounding interactively. Using the software, tumor volume is precisely calculated.
Hintergrund: Das Glioblastom ist ein hoch aggressiver und maligne Hirntumor, welcher schwer zu resektieren ist. Der Erfolg der operativen Entfernung hat einen direkten Einfluss auf die Überlebenszeit des Patienten. Um das Ausmass der Resektion festzustellen, wird das prä- und postoperative Tumorvolumen mithilfe von Magnetresonanztomografie (MRT)-Aufnahmen berechnet und verglichen. Materialien und Methoden: Eine aktive Kontur wurde zur Segmentierung von Glioblastom Hirntumoren auf präoperativen kontrastverstärkten T1-gewichteten MRTAufnahmen implementiert. Die selbstentwickelte Software erlaubt die Segmentierung auf axialen, koronalen und sagittalen MRT-Aufnahmen. Das Tumorvolumen wurde von den segmentierten Tumorkonturen mittels Delaunay-Triangulation berechnet und dargestellt. Um die Segmentierung, Tumordarstellung und Volumenberechnung zu erleichtern, wurde eine grafische Benutzeroberfläche in MATLAB (Version 7.2) entwickelt. Zwei MRT Datensätze von Glioblastom-Patienten wurden verwendet und die Wiederhol- und die Reproduzierbarkeit der Volumenberechnung wurden getestet. Ergebnisse: Die Tumorvolumina für den Datensatz 1 und Datensatz 2 betragen 62,7 bzw. 39,0 cm3. Die Segmentierung der Tumore durch verschiedene Benutzer (n=4) lieferte ein Volumen von 62,5 ± 0,3 und 42,6 ± 3,5 cm3. Die Segmentierung des zweiten Datensatzes verursachte Probleme wie Untersegmentierung durch Cerebrospinalflüssigkeit oder den Schädel sowie Übersegmentierung durch schwacheTumorkonturen. Schlussfolgerung: Aktive Konturen sind in der Lage Hirntumore korrekt zu segmentieren. Die Volumenberechnung und -darstellung erlaubt dem Benutzer, den Tumor, sein Gewebe und das umliegende Hirngewebe, interaktiv zu sondieren. Durch die Verwendung der Software wird das Tumorvolumen präzise berechnet.

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.

A brain tumor diagnosis is traumatic and has a devastating impact upon the caregiver and the family unit. The effects of the tumor growth and treatment often cause significant neurologic injury and dramatically affect the quality of life (QOL) for the patient and their entire family unit. Caregivers are constantly challenged to provide care, yet they feel untrained and underprepared as they struggle to adjust to new roles and responsibilities. The purpose of this study is to gain an understanding of the lived experiences of caregivers of individuals with primary benign brain tumor (PBBT). An interpretive phenomenological analysis approach was used to explore the experiences of 10 caregivers. Bowen's family systems theory provided an understanding of how families respond to changes in their family system resulting from a member of the family having a PBBT. A nonprobability sampling technique was used to recruit participants from 2 virtual support groups. Data were collected through semistructured interviews guided by an interview template. Interviews were transcribed and analyzed following the Smith tradition of inquiry until data saturation was reached. Three major themes emerged from the data: experiencing new challenges, responding to initial diagnosis, and facing challenges with family and friends. Caregivers experience a wide variety of responsibilities that are physically and psychologically challenging, which can negatively affect the QOL for the caregiver and the patient. These findings can be used by healthcare providers to identify resources to alleviate the unanticipated demands caregivers experience. Future studies are needed to explore how best to decrease challenges experienced by caregivers of individuals with PBBT.

Proton radiation therapy is a form of external radiation that uses charged particles which have distinct physical advantages to deliver the majority of its dose in the target while minimizing the dose of radiation to normal tissues. In children who are particularly susceptible even at low and medium doses of radiation, the significant reduction of integral dose can potentially mitigate the incidence of side effects and improve quality of life. The aim of the first part of the thesis is to describe the physical and radiobiological properties of protons, the Proton Therapy Center of Trento (TCPT) active for clinical purpose since 2014, which use the most recent technique called active pencil beam scanning. The second part of the thesis describes the preliminary clinical results of 23 pediatric patients with central nervous system tumors as well as of two aggressive pediatric meningiomas treated with pencil beam scanning. All the patients were particularly well-suited candidates for proton therapy (PT) for possible benefits in terms of survival and incidence of acute and late side effects. We reported also a multicentric experience of 27 medulloblastoma patients (median age 6 years, M/F ratio 13/14) treated between 2015 and 2020 at TPTC coming from three Pediatric oncology centers: Bologna, Florence, and Ljubljana, with a focus on clinical results and toxicities related to radiotherapy (RT). Proton therapy was associated with mostly mild acute and late adverse effects and no cases of CNS necrosis or high grade of neuroradiological abnormailities. Comparable rates of survival and local control were obtained to those achievable with conventional RT. Finally, we performed a systematic review to specifically address the safety of PT for pediatric CNS patients, late side effects and clinical effectiveness after PT in this patient group.

Brain computer interfaces (BCIs) are systems that allow users to interact with devices without relying on the neuromuscular pathways. This interaction is achieved by allowing the system to read the electrical activity of the brain and teaching it to map certain patterns of activation to certain commands. There are many applications for BCIs ranging from controlling prosthetics to gaming, but adapting both the user and the system to one another is a time and resource consuming process. Even more problematic, BCIs tend to only perform well for a single user and only for a limited time. This paper aims to investigate the accuracy of single-subject singlesession BCIs on other subjects and other sessions. To that end three different classifiers, a Support Vector Machine (SVM), Convolutional Neural Network (CNN) and Long Short-Term Memory network (LSTM) are developed and tested on a data set consisting of five subjects, two sessions for a binary classification task. Our results show that training on single-subject single-session data leads to an average cross-subject accuracy of 45-50% and an average cross-session accuracy of 50-55%. We find that there is no statistically significant difference in accuracy depending on the classifier used and discuss factors that affect generalization such as model complexity and good subjects.
Brain computer interfaces (BCIs) är system som gör det möjligt för användare att interagera med apparater utan behov av de neuromuskulära banorna. Den här interaktionen möjliggörs genom att systemet läser den elektriska aktiviteten i hjärnan och lär sig associera vissa mönster av aktivitet till vissa kommandon. Det finns många användningsområden för BCIs, från att kontrollera proteser till spel, men att anpassa både användaren och systemet till varandra är en process som kräver både tid och resurser. Än värre, BCIs tenderar att bara funka bra för en enskild användare och bara under en begränsad tid. Den här rapporten avser undersöka hur bra ett BCI system tränat på data för ett subjekt och en session är på klassificering av data för andra subjekt och andra sessioner. Tre typer av klassificerare, en Support Vector Machine (SVM), Convolutional Neural Network (CNN) och Long Short-Term Memory network (LSTM) byggs och utvärderas på data från fem subjekt över två sessioner på en binär klassificeringuppgift. Våra resultat indikerar att träning på data för ett subjekt, en session leder till en genomsnittlig pricksäkerhet på 45-50% på andra subjekt, 50-55% på andra sessioner. Vi finner även att det inte finns någon statistiskt signifikant skillnad i pricksäkerhet beroende på vilken typ av klassificerare som används och diskuterar faktorer som påverkar generalisering såsom modellkomplexitet och bra subjekt.

This dissertation focuses on the development of three classes of brain-inspired machine learning classification models. The models attempt to emulate (a) multi-sensory integration, (b) context-integration, and (c) visual information processing in the brain.The multi-sensory integration models are aimed at enhancing object classification through the integration of semantically congruent unimodal stimuli. Two multimodal classification models are introduced: the feature integrating (FI) model and the decision integrating (DI) model. The FI model, inspired by multisensory integration in the subcortical superior colliculus, combines unimodal features which are subsequently classified by a multimodal classifier. The DI model, inspired by integration in primary cortical areas, classifies unimodal stimuli independently using unimodal classifiers and classifies the combined decisions using a multimodal classifier. The multimodal classifier models are be implemented using multilayer perceptrons and multivariate statistical classifiers. Experiments involving the classification of noisy and attenuated auditory and visual representations of ten digits are designed to demonstrate the properties of the multimodal classifiers and to compare the performances of multimodal and unimodal classifiers. The experimental results show that the multimodal classification systems exhibit an important aspect of the “inverse effectiveness principle” by yielding significantly higher classification accuracies when compared with those of the unimodal classifiers. Furthermore, the flexibility offered by the generalized models enables the simulations and evaluations of various combinations of multimodal stimuli and classifiers under varying uncertainty conditions. The context-integrating model emulates the brain’s ability to use contextual information to uniquely resolve the interpretation of ambiguous stimuli. A deep learning neural network classification model that emulates this ability by integrating weighted bidirectional context into the classification process is introduced. The model, referred to as the CINET, is implemented using a convolution neural network (CNN), which is shown to be ideal for combining target and context stimuli and for extracting coupled target-context features. The CINET parameters can be manipulated to simulate congruent and incongruent context environments and to manipulate target-context stimuli relationships. The formulation of the CINET is quite general; consequently, it is not restricted to stimuli in any particular sensory modality nor to the dimensionality of the stimuli. A broad range of experiments are designed to demonstrate the effectiveness of the CINET in resolving ambiguous visual stimuli and in improving the classification of non-ambiguous visual stimuli in various contextual environments. The fact that the performance improves through the inclusion of context can be exploited to design robust brain-inspired machine learning algorithms. It is interesting to note that the CINET is a classification model that is inspired by a combination of brain’s ability to integrate contextual information and the CNN, which is inspired by the hierarchical processing of visual information in the visual cortex. A convolution neural network (CNN) model, inspired by the hierarchical processing of visual information in the brain, is introduced to fuse information from an ensemble of multi-axial sensors in order to classify strikes such as boxing punches and taekwondo kicks in combat sports. Although CNNs are not an obvious choice for non-array data nor for signals with non-linear variations, it will be shown that CNN models can effectively classify multi-axial multi-sensor signals. Experiments involving the classification of three-axis accelerometer and three-axes gyroscope signals measuring boxing punches and taekwondo kicks showed that the performance of the fusion classifiers were significantly superior to the uni-axial classifiers. Interestingly, the classification accuracies of the CNN fusion classifiers were significantly higher than those of the DTW fusion classifiers. Through training with representative signals and the local feature extraction property, the CNNs tend to be invariant to the latency shifts and non-linear variations. Moreover, by increasing the number of network layers and the training set, the CNN classifiers offer the potential for even better performance as well as the ability to handle a larger number of classes. Finally, due to the generalized formulations, the classifier models can be easily adapted to classify multi-dimensional signals of multiple sensors in various other applications.

Regulation of cancer cell phenotype by the tumor microenvironment has motivated further investigation into how microenvironmental factors could contribute to tumor initiation, development, and therapeutic resistance. Analyzing how the microenvironment drives tumor development and cancer cell heterogeneity is particularly important in cancers such as glioblastoma multiforme (GBM) that have no known risk factors and are characterized by a high degree of heterogeneity. GBM patients have a median survival of 15 months and therefore are in great need of more effective therapeutic options. The goal of this research is to generate in vitro models of the heterogeneous brain tumor microenvironment, with a focus on vascular dynamics, to probe the impact of microenvironmental cues on tumor progression and to integrate the tumor models with highly sensitive analytical tools to characterize the epigenome of discrete phenotypic subpopulations that contribute to intratumoral cellular heterogeneity. As GBM tumors are characterized by a dense vasculature, we delved into microenvironmental factors that may be promoting angiogenesis. The correlations emerging between inflammation and cancer led to analysis of the inflammatory molecule lipopolysaccharide (LPS). We utilized 3D micro-tissue models to simulate vascular exposure to ultra-low chronic inflammatory levels of LPS and observed an increase in vascular formation when brain endothelial cells were exposed to ultra-low doses of LPS. We also utilized our micro-tissue models to analyze histone methylation changes across the epigenome in response to microenvironmental cues, namely culture dimensionality and oxygen status. The H3K4me3 modification we analyzed is associated with increased gene transcription, therefore the alterations we observed in H3K4me3 binding across the genome could be a mechanism by which the tumor microenvironment is regulating cancer cell phenotype. Lastly, we developed a microfluidic platform in which vascular dynamics along with microenvironmental heterogeneities can be modeled in a more physiologically relevant context. We believe the studies presented in this dissertation provide insight into how vasculature primed by chronic inflammation and epigenetic alterations in tumor cells could both contribute to enhanced tumor development. Modeling these biological processes in our advanced microfluidic platform further enables us to better understand microenvironmental regulation of tumor progression, uncovering new potential therapeutic targets.
PHD

Brain and central nervous cancer presents a significant clinical burden, accounting for 2.4% of all cancer deaths. High grade glioma is particularly deadly, with 5 year survival times of 35% or less. Traditional treatment includes tumor resection followed by radiation therapy or chemotherapy. Aggressive resection is essential in order to prolong patient life. In fact, several studies have shown that life expectancy increases with increased extent of resection. Extent of resection is burdened by the fact that surgeons must be careful not to remove functional brain tissue. Resection is incomplete more often than not due to lack of visual cues for the surgeon. He must rely on tactile sensation to distinguish tumor from healthy tissue. Methods such as intraoperative MRI and CT exist, but these require expensive equipment and special training that is not available in all surgical environments. Some laboratories have proposed small molecule dyes to solve this problem, but these are insufficient when used in an invasive tumor model. It was the goal of this research to provide an objective cue in the form of a nanoencapsulated visible dye without the need for additional equipment of changes to the surgery process itself other than injection of the dye. We hypothesized that the nanocarrier would allow staining of the tumor through passive targeting by taking advantage of the enhanced permeability and retention effect. Once the nanocarriers have reached the desired target, they would not diffuse out into healthy tissue due to their large size compared to small molecule dyes, which readily diffuse out and stain healthy tissue. To test this hypothesis, we prepared and characterized a liposomal nanocarrier encapsulating Evans blue dye. The nanocarrier was tested for safety in vitro and in vivo, then used to delineate tumor margins in an invasive rat glioma model in vivo. Microscopic analysis was then conducted to ensure only tumor tissue was stained by the nanocarrier. This thesis presents a successful method of tumor border delineation to provide surgeons with positive visual cues without the need for changes in surgical environment or techniques.

Change in adaptive outcomes of children who are treated for brain tumors is examined using longitudinal data. The children received different types of treatment from none to any combinations of three treatments, which are surgery, radiation and chemotherapy. In this thesis, we use mixed model to find the significant variables that predict change in outcomes of communication skill, daily living skills and socialization skill. Fractional polynomial transformation method and Gompertz method are applied to build non-linear longitudinal curves. We use PRESS as the criterion to compare these two methods. Comparison analysis shows the effect of each significant variable on adaptive behaviors over time. In most cases, model with Gompertz method is better than that with Transformation method. Significant predictors of change in adaptive outcomes include Time, Gender, Surgery, SES classes, interaction between Time and Radiation, interaction between Time and Gender, interaction between Age and Gender.

碩士
國立臺灣海洋大學
資訊工程學系
104
Magnetic resonance imaging (MRI) have recently played an important role in clinical research and quantitative analysis for medical doctor to study the structural features and the functional characteristics of the internal body parts. There is a practical limit on the signal-to-noise ratio (SNR) when acquiring MR image data. Therefore, it is important to improve the SNR of the MR images used during quantitative analysis. This is usually done by using a denoising method. We proposed a wavelet-based bilateral filter method, called Multi-Bilateral Filter (MBF) in the first part of this thesis. The experiment results show the proposed scheme is comparable to the traditional bilateral methods. The Computer-Aided Diagnosis (CAD) in detecting brain disease is progressively important in clinic diagnosis. The automatically classifying system can help physician analyzing and predicting the normal or tumor brains with magnetic resonance imaging. Secondly, we propose a brain tumor diagnosis algorithm with T2-weighted magnetic resonance images of brain. The detecting procedure includes the denoising method with MBF, feature extraction from space and wavelet domains, feature reduction with principal component analysis and final classification with support vector machine learning. In comparison, classification accuracy with a best way is obtained by the proposed method based on with cutting white matter and combined features with spatial Gary Level Co-occurrence Matrix, wavelet variance and Besov spaces texture characterization.

The brain tumor is a very common disease among humans, it can be deadly if not diagnosed in its early stage. Magnetic Resonance Imaging (MRI) is commonly used technique to diagnose cancer but despite providing highly valuable information regarding tumor, it also prone to give human error. Recently, various Computer-Aided Diagnosis (CAD) techniques are being developed to improve performance of MRI. Computer vision systems make this process automatic and more accurate detection so that diagnosis can be done properly. Here presents a general Metaheuristic algorithm followed by review on different Metaheuristic Optimization Techniques based CAD systems which provides better detection and classification results for MR Images. The word Meta means “trial and run” and heuristic means an “approach” to find a solution, so together metaheuristic is a method that uses a trial and run approach to find an effective solution. Our main aim here to detect the tumor type whether it is in benign stage or in malignant stage. This is done basically with help of mainly 4 steps namely Enhancement- which is particularly removing the noise, Segmentation – which makes the effected region visible , Feature Extraction/selection- which helps to extract the features which helps in classifying the tumor and finally Classification-which identifies the type of tumor.

MRI images [8] play a significant influence in brain tumor classification and detection but instead of having detection and classification using the medical equipment which is a radiologists or clinical professionals do a time-consuming and laborious task where accuracy depends only on the experience only, it can be beneficial to detect and classify the brain tumor by deep learning techniques and algorithms. As a result, using computer-assisted technologies to circumvent these limits is becoming increasingly vital. In this paper, the early detected and diagnosed brain tumor images along with their csv data has been used to find out the accuracy of the CNN algorithm for tumor detection and SVM algorithm for tumor classification into benign and malignant. HOG has been used for the feature extraction. After performing the experiment, it was observed that CNN achieved the detection accuracy of 87.02% and further tumor classification by employing SVM, the highest accuracy achieved was 96.35%. The experiment proved a very good accuracy of detection and classification even after using three different methods in the whole procedure.

Dissertação de mestrado integrado em Engenharia Biomédica (área de especialização em Informática Médica)
Prognosis and patient stratification for brain tumors is an important and clinically relevant task and a precise treatment outcome prediction would allow to choose an adequate therapy strategy and schedule the most appropriate follow-up examinations. Magnetic Resonance Imaging (MRI) is an already know imaging technique to assess these tumors. Next to medical imaging, other clinical information is important for patient management, e.g. genetic markers like O6-Methyl-Guanine-Methyl-Transferase (MGMT) methylation is a well-known prognostic marker in Glioblastoma (GBM) tumors. Therefore, the main goal of this thesis was to study Deep Learning (DL) approaches to combine MRI with non-image clinical data in two different classification scenarios: brain tumor segmentation and patient outcome prediction. There are studies that combine these two types of data, however, in two steps: extracting MRI features and then combining them with relevant non-image data. Here, end-to-end DL architectures with two input layers are presented, as well as an infrastructure that allows the easy development of future Machine Learning (ML) /DL models that consumes these two types of data in a clinical context. In this way, the classification in both scenarios is done in a single step, where Convolution Layers perform the feature extraction in MRI input. In brain tumor segmentation, the model with combined data achieved a slightly better Dice Similarity Coefficient (DSC) (0.894 ± 0.025) over image only model (0.882 ± 0.025). As for patient outcome prediction, when trying to predict the Progression-Free Survival (PFS) class (“bad”,” medium” and “good” outcomes), the combined model didn't improve when compared with the model where only MRI was used. Both models, however, outperformed models where only non-image data was used. The segmentation results point to a positive influence when adding the clinical information to MRI. Nevertheless, there is a lot more to investigate in this field, not only in the model architecture, but also in selecting relevant clinical information. In same way, more tests should be run for patient outcome prediction, especially using Overall Survival (OS) information.
O prognóstico de pacientes com tumores cerebrais é uma tarefa importante e clinicamente relevante. Uma previsão precisa do resultado do(s) tratamento(s) permitiria escolher uma estratégia de terapia adequada e agendar os exames de seguimento mais apropriados. A Ressonância Magnética (RM) é uma técnica de imagem já conhecida na avaliação deste tipo de tumores. Outras informações clínicas como o marcador tumoral MGMT, são também relevantes no prognóstico de tumores de Glioblastoma (GBM). Neste sentido, o objetivo principal desta tese foi estudar modelos de Deep Learning (DL) para combinar imagens de RM com informação clínica e aplicá-los em dois cenários de classificação: segmentação de tumores cerebrais e previsão do prognóstico do paciente. Existem estudos que combinam estes dois tipos de dados, porém, em duas fases: extraindo atributos das imagens de RM e combinando-as posteriormente com informação clínica relevante. Contrariamente, aqui são apresentadas arquiteturas end-to-end de DL com duas camadas de entrada, assim como uma infraestrutura que permite um fácil desenvolvimento de modelos de Machine Learning (ML) / DL capazes de consumir estes dois tipos de dados num ambiente hospitalar. Desta forma, a classificação em ambos os cenários é feita em um único passo, onde camadas de convolução realizam a extração de características das imagens na entrada de MRI. No cenário de segmentação de tumores, o modelo que utilizou os dados combinados obteve um DSC (Dice Similarity Coefficient) de 0.894 ± 0.025, superando ligeiramente o modelo que usou apenas imagens de RM (0.882 ± 0.025). Quanto à previsão do prognóstico do paciente, classificando com a medida de Progression-Free Survival (PFS) (“bad”,” medium” e “good”) o modelo combinado não melhorou quando comparado com o modelo que apenas utilizou imagens de RM. Ambos os modelos, no entanto, superaram modelos que utilizaram apenas os dados clínicos. Os resultados da segmentação apontam para uma influência positiva na adição de informação clínica às imagens de RM. No entanto, ainda há muito a investigar neste campo, não apenas na arquitetura do modelo, mas também na seleção de informações clínicas relevantes. Da mesma forma, mais testes deverão ser executados para a previsão do prognóstico do paciente, especialmente usando a medida de Overall Survival (OS).

Magnetic resonance (MR) imaging has been playing an important role in neuroscience research for studying brain images where MR’s soft tissue contrast and non invasiveness are clear advantages. MR images can also be used to determine normal and abnormal types of brain. Moreover, the MRI characteristics will help the doctor to avoid the human error in manual interpretation of medical content. Computer-based classification has remained largely experimental work with approaches. Here a work is done by simulating a method in Matlab using artificial neural network to automatically classify brain MRI images. The diagnosis method consists of three stages firstly feature extraction using discrete wavelet transforms. Wavelets seem to be a suitable tool for this task, because they allow analysis of images at various levels of resolution. Then the features are reduced using principal component analysis (PCA). In the last stage artificial neural network (ANN) is used as a multi class classification technique to classify between normal & brain tumor infected MRI Images & also classify different brain tumor images according to the different location of Tumor in the brain. We obtain good classification rate with the less number of features. The results show that the method is robust and effective.

碩士
國立陽明大學
衛生資訊與決策研究所
92
There are over 120 different types of brain tumors, making effective treatment very complicated. Classification of brain tumors accurately can not only help the doctors to treat the patients correctly but also help doctors to do research and teaching in this field efficiently. The objective of our study was to classify pathological reports into different classes of brain tumors automatically according to World Health Organization 2000 classification of brain tumors. We developed pattern-matching rules called Brain-Tumor Classifier processing pathological reports and classifying brain tumors automatically. We compared Brain-Tumor Classifier against a gold standard that was established by three experts judging 276 records. In this testing set, Brain-Tumor Classifier had a specificity of 99.74% (versus 99.79 ~ 99.9 % for the physicians), a positive predictive value of 91.67% (versus 82.35 ~ 94.92 % for the physicians) while maintaining a reasonable sensitivity of 90.83% (versus 85.91 ~ 97.93 % for the physicians). In addition, it had accuracy of 91.1%. We conclude that automatic processing of pathological reports for classification of brain tumors is feasible and useful.