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.

A possibilidade de recuperar e comparar imagens usando as suas características visuais intrínsecas é um recurso valioso para responder a consultas por similaridade em imagens médicas. Desse modo, a agregação desses recursos aos Sistemas de Arquivamento e Comunicação de Imagens (Picture Archiving and Communication Systems - PACS) vêm potencializar a utilidade e importância destes no contexto de atividades tais como ensino e treinamento de novos radiologistas, estudos de casos e auxílio ao diagnóstico de forma geral, uma vez que as consultas por similaridade permitem que casos parecidos possam ser facilmente recuperados. O trabalho apresentado nesta tese possui duas vertentes. Primeiro, ele apresenta novos métodos de extração e de características, com o objetivo de obter a essência das imagens, considerando um critério específico. Os atributos obtidos pelos algoritmos de extração são armazenados em vetores de características para posteriormente serem utilizados para indexar e recuperar as imagens baseando-se em seu conteúdo, para responder a consultas por similaridade. Há uma relação próxima entre os vetores de características e as funções de distância utilizadas para compará-los. Assim, a segunda parte deste trabalho trata da proposta, análise e comparação de novas famílias de funções de distância. As funções de distância propostas têm por objetivo tratar o problema do gap semântico, o qual representa o principal obstáculo das funções de distância tradicionais, derivadas da família Lp, quando processam consultas por similaridade. As principais contribuições desta tese incluem o desenvolvimento de novos métodos de extração e comparação de características de imagens, que operam sobre os três principais descritores de baixo nível de imagens: distribuição de cor, textura e forma. Os experimentos realizados mostraram que os ganhos em precisão são maiores para os métodos propostos, quando comparados com algoritmos tradicionais. No que diz respeito às famílias de funções de distância propostas (WAID e SAID), pelos resultados iniciais obtidos, podemos afirmar que eles são bastante promissores no sentido de se aproximarem da expectativa do usuário, no momento de comparar imagens. Os resultados obtidos com esse trabalho podem ser futuramente integrados aos PACS. Particularmente, pretendemos acrescentar novos algoritmos e métodos ao cbPACS, que consiste em um sistema PACS em construção, desenvolvido em uma colaboração entre o Grupo de Bases de Dados e Imagens (GBDI) do Instituto de Ciências Matemáticas e de Computação - USP e o Centro de Ciências da Imagens e Física Médica (CCIFM) da Faculdade de Medicina de Ribeirão Preto - USP / The ability of retrieving and comparing images using their inherent pictorial information is a valuable asset to answer similarity queries over medical images. Thus, having such resources added in Picture Archiving and Communication Systems (PACS) increase their applicability and importance in the context of teaching and training new radiologists on diagnosing, since that similar cases can be easily retrieved. Similarity queries also play an important role on gathering close images, what allows to perform case studies, as well as to aid on diagnosing. The work presented in this thesis is twofold. First, it presents new feature extraction techniques, which aim at obtaining the essence of the images regarding a given criteria. The features obtained by the algorithms are stored in feature vectors and employed to index and retrieve the images by content, in order to answer similarity queries. There is a close relationship among feature vectors and the distance function employed to compare them. Thus, the second, part of this work concerns the comparison, analysis and proposal of new families of distance functions to compare the features extracted from the images. The distance functions proposed intend to deal with the semantic gap problem, which is the main drawback of the traditional distance functions derived from the Lp metrics when processing similarity queries. The main contributions of this thesis include the development of new image feature extractors that works on the three aspects of raw image data (color distribution, texture and shape). The experiments have shown that the gain in precision are higher for all the feature extractors proposed, when comparing with the state-of-the-art algorithms. Regarding the two families of distance functions WAID and SAID proposed, by the initial experiments performed we can claim that they are very promising on preserving the user expectation when comparing images. The results provided by this work can be straightforwardly integrated to PACS. Particularly, we intend to add the new algorithms and methods to cbPACS, which is under joined development between the Image Data Base Group of Instituto de CiLncias Matemáticas e de Computaçno of USP and Centro de CiLncias de Imagens e Física Médica of Faculdade de Medicina de Ribeirno Preto of USP

El texto completo de este trabajo no está disponible en el Repositorio Académico UPC por restricciones de la casa editorial donde ha sido publicado. / This paper proposes a technological solution using a predictive analysis model to identify and reduce the level of risk for type 2 diabetes mellitus (T2DM) through a wearable device. Our proposal is based on previous models that use the auto-classification algorithm together with the addition of new risk factors, which provide a greater contribution to the results of the presumptive diagnosis of the user who wants to check his level of risk. The purpose is the primary prevention of type 2 diabetes mellitus by a non-invasive method composed of the phases: (1) Capture and storage of risk factors; (2) Predictive analysis model; (3) Presumptive results and recommendations; and (4) Preventive treatment. The main contribution is in the development of the proposed application. / Revisión por pares

Ultrasound imaging is one of the fundamental examinations of thyroid nodules, determining whether a patient undergoes a cytological examination, which is essential for the decision on a possible thyroid surgery. Unfortunately, the cytological examination has limited specificity and potential surgery carries risks. Therefore, other diagnostic methods are being sought with hope that they will be able to bring more certainty into diagnostics. One of the new methods is computer-aided diagnosis (CAD), which exhibits promising results using image analysis and machine learning. In this study, we present two somewhat similar, yet different, CAD approaches. The first approach is based on analysing entire nodules using a Segmentation Based Fractal Texture Analysis (SFTA) algorithm that splits the image into individual grayscale bands. Using this approach, we have achieved an accuracy of 92.4% using random forests (RF) and 95% using support vector machines (SVM) on a data set of 40 images evaluated by the cross-validation method. The second CAD approach is also based on the method of multiple image thresholding, but the difference is, that a larger number of predictors describing the binary texture are extracted from the individual grayscale bands. Furthermore, the analysis did not take place on whole nodules, but on...

Breast cancer is one of the most common cancers among women in the world, and the average error rate among radiologists during diagnosis is 30%. Computer-aided medical diagnosis aims to assist doctors by giving them a second opinion, thus decreasing the error rate. Convolutional neural networks (CNNs) have shown to be good for visual detection and recognition tasks, and have been explored in combination with transfer learning. However, the performance of a deep learning model does not only rely on the model itself, but on the nature of the dataset as well In breast cancer diagnosis, the area surrounding a mass provides useful context for diagnosis. In this study, we explore providing different amounts of context to the CNN model ResNet50, to see how it affects the model’s performance. We test masses with no additional context, twice the amount of original context and four times the amount of original context, using 10-fold cross-validation with ROC AUC and average precision (AP ) as our metrics. The results suggest that providing additional context does improve the model’s performance. However, giving two and four times the amount of context seems to give similar performance. / Bröstcancer är en av de vanligaste cancersjukdomar bland kvinnor i världen, och den genomsnittliga felfrekvensen under diagnoser är 30%. Datorstödd medicinsk diagnos syftar till att hjälpa läkare genom att ge dem en andra åsikt, vilket minskar felfrekvensen. Konvolutionella neurala nätverk (CNNs) har visat sig vara bra för visuell detektering och igenkännande, och har utforskats i samband med det s.k. “transfer learning”. Prestationen av en djup inlärningsmodell är däremot inte enbart beroende på modellen utan också på datasetets natur. I bröstcancerdiagnos ger området runt en bröstmassa användbar kontext för diagnos. I den här studien testar vi att ge olika mängder kontext till CNNmodellen ResNet50, för att se hur det påverkar modellens prestanda. Vi testar bröstmassor utan ytterligare kontext, dubbelt så mycket som den originala mängden kontext och fyra gånger så mycket som den orginala mängden kontext, med hjälp av “10-fold cross-validation” med ROC AUC och “average precision” (AP ) som våra mätvärden. Resultaten visar att mer kontext förbättrar modellens prestanda. Däremot verkar att ge två och fyra gånger så mycket kontext resultera i liknande prestanda.

One of the most globally common types of cancer is skin cancer, where melanoma is the most deadly form. An important and promising tool for diagnosing diseases such as skin cancer is computer aided diagnostics, a tool which utilizes machine learning to predict and classify cancer. Limiting the complexity of the data, known as feature selection, can potentially improve classification accuracy. This report evaluates the accuracy of four different classifiers - Support Vector Machine, Naive Bayes, Decision Tree and Artificial Neural Network - with four different feature selection methods - Sequantial Forward Selection, Sequantial Backward Selection, Entropy and Principal Component Analysis - on the PH2 skin cancer dataset, containing dermoscopic images of skin lesions and their respective metadata. The findings reveal that all feature selection methods led to an improved accuracy rate on at least one classifier compared to not using feature selection. Furthermore, certain feature selection methods resulted in a significant gain in accuracy, indicating the potential value of feature selection techniques in improving the accuracy and efficiency of machine learning classifiers in computer-aided diagnosis systems for melanoma skin cancer detection. However, the results also underscore the importance of careful selection of the number of features to avoid adverse effects on model performance. This research contributes to the field by demonstrating the impact of feature selection methods on melanoma skin cancer detection and highlighting considerations for their application. / En av de globalt vanligaste typerna av cancer är hudcancer, där melanom är den mest dödliga typen. Ett viktigt och effektivt verktyg för att diagnostisera sjukdomar som hudcancer är datorstödd diagnostik, ett verktyg som använder maskininlärning för att förutse och klassificera cancer. Att begränsa komplexiteten i data, känt som attributurval, kan potentiellt förbättra klassificeringsnoggrannheten. Denna rapport utvärderar noggrannheten hos fyra olika klassificerare - ”Support Vector Machine”, ”Naive Bayes”, ”Decision Tree” och ”Artificial Neural Network” - med fyra olika attributurvalsmetoder - ”Sequantial Forward Selection”, ”Sequantial Backward Selection”, ”Entropy” and ”Principal Component Analysis” - på PH2 hudcancerdatasetet, som innehåller dermoskopiska bilder av hudlesioner och deras respektive metadata. Resultaten visar att alla attributurvalsmetoder ledde till en förbättrad noggrannhetsgrad på minst en klassificerare jämfört med att inte använda attributurval. Dessutom resulterade vissa attributurvalsmetoder i en betydande ökning i noggrannhet, vilket indikerar det potentiella värdet av attributurvalstekniker för att förbättra noggrannheten och effektiviteten hos maskininlärningsklassificerare i datorstödda diagnossystem för detektering av melanom hudcancer. Däremot understryker resultaten också vikten av noggrant urval av antalet attribut för att undvika negativa effekter på modellens prestanda. Denna forskning bidrar till fältet genom att demonstrera inverkan av attributurvalsmetoder på detektering av melanom hudcancer och belysa överväganden för deras tillämpning.

Alzheimer’s disease is a neurodegenerative disease that affects approximately 6% of the global population aged over 65 and is forecasted to become even more prevalent in the future. Accurately diagnosing the disease in an early stage can play a large role in improving the quality of life for the patient. One key development for performing this diagnosis is applying machine learning to perform computer-aided diagnosis. Current research in the field has been focused on removing assumptions about the used data sets, but in doing so they have often discarded objective metadata such as the patient’s age, sex or priormedical history. This study aimed to investigate the effect of including such metadata as additional input features to neural networks used for diagnosing Alzheimer’s disease through binary classification of magnetic resonance imaging scans. Two similar neural networks were developed and compared, one with these additional features and the other without them. Including the metadata led to significant improvements in the network’s classification accuracy, and should therefore be considered in future computer-aided diagnostic systems for Alzheimer’s disease. / Alzheimers sjukdom är en form av demens som påverkar ungefär 6% av den globala befolkningen som är äldre än 65 och förutspås bli ännu vanligare i framtiden. Tidig diagnos av sjukdomen är viktigt för att säkerställa högre livskvalitet för patienten. En viktig utveckling inom fältet är datorstödd diagnos av sjukdomen med hjälp av maskininlärning. Dagens forskning fokuserar på att ta bort subjektiva antaganden om datamängden som används, men har ofta även förkastat objektiv metadata såsom patientens ålder, kön eller tidigare medicinska historia. Denna studier ämnade därför undersöka om inkluderandet av denna metadata ledde till bättre prestanda hos neuronnät som används för datorstödd diagnos av Alzheimers genom binär klassificering av bilder tagna med magnetisk resonanstomografi. Två snarlika neuronnät utvecklades och jämfördes, med skillnaden att den ena även tog metadata om patienten som indata. Inkluderandet av metadatan ledde till en markant ökning i neuronnätets prestanda, och bör därför övervägas i framtida system för datorstödd diagnos av Alzheimers sjukdom.

No description available.

Le cancer de la prostate est la deuxième cause de mortalité chez l’homme en France. L’IRM multiparamétrique est considérée comme la technique la plus prometteuse pour permettre une cartographie du cancer, ouvrant la voie au traitement focal, alternatif à la prostatectomie radicale. Néanmoins, elle reste difficile à interpréter et est sujette à une forte variabilité inter- et intra-expert, d’où la nécessité de développer des systèmes experts capables d’aider le radiologue dans son diagnostic. Nous proposons un système original d’aide au diagnostic (CAD) offrant un second avis au radiologue sur des zones suspectes pointées sur l’image. Nous évaluons notre système en nous appuyant sur une base de données clinique de 30 patients, annotées de manière fiable et exhaustive grâce à l’analyse des coupes histologiques obtenues par prostatectomie. Les performances mesurées dans des conditions cliniques auprès de 12 radiologues, sans et avec notre outil, démontrent l’apport significatif de ce CAD sur la qualité du diagnostic, la confiance des radiologues et la variabilité inter-expert. La création d’une base de corrélations anatomo-radiologiques est une tâche complexe et fastidieuse. Beaucoup d’études n’ont pas d’autre choix que de s’appuyer sur l’analyse subjective d’un radiologue expert, entâchée d’incertitude. Nous proposons un nouveau schéma de classification, basé sur l’algorithme du séparateur à vaste marge (SVM), capable d’intégrer, dans la fonction d’apprentissage, l’incertitude sur l’appartenance à une classe (ex. sain/malin) de certains échantillons de la base d’entraînement. Les résultats obtenus, tant sur des exemples simulés que sur notre base de données cliniques, démontrent le potentiel de ce nouvel algorithme, en particulier pour les applications CAD, mais aussi de manière plus générale pour toute application de machine learning s’appuyant sur un étiquetage quantitatif des données / Prostate cancer is one of the leading cause of death in France. Multi-parametric MRI is considered the most promising technique for cancer visualisation, opening the way to focal treatments as an alternative to prostatectomy. Nevertheless, its interpretation remains difficult and subject to inter- and intra-observer variability, which motivates the development of expert systems to assist radiologists in making their diagnosis. We propose an original computer-aided diagnosis system returning a malignancy score to any suspicious region outlined on MR images, which can be used as a second view by radiologists. The CAD performances are evaluated based on a clinical database of 30 patients, exhaustively and reliably annotated thanks to the histological ground truth obtained via prostatectomy. Finally, we demonstrate the influence of this system in clinical condition based on a ROC analysis involving 12 radiologists, and show a significant increase of diagnostic accuracy, rating confidence and a decrease in inter-expert variability. Building an anatomo-radiological correlation database is a complex and fastidious task, so that numerous studies base their evaluation analysis on the expertise of one experienced radiologist, which is thus doomed to contain uncertainties. We propose a new classification scheme, based on the support vector machine (SVM) algorithm, which is able to account for uncertain data during the learning step. The results obtained, both on toy examples and on our clinical database, demonstrate the potential of this new approach that can be extended to any machine learning problem relying on a probabilitic labelled dataset

A análise de imagens médicas auxiliada por computador permite a análise quantitativa das anormalidades e garante maior precisão diagnóstica. Esse tipo de análise é importante para medicina nuclear com Single Photon Emission Computed Tomography (SPECT), pois no grupo de dados tridimensionais de imagens, padrões sutis de anormalidades muitas vezes são importantes achados clínicos. Porém, as imagens podem sofrer interferência de artefatos de atenuação da emissão de fótons por partes moles corporais, o que reduz sua acurácia diagnóstica. Desde que se possuam parâmetros de atenuação computados em um modelo que permita a comparação com imagens de um dado paciente, a interferência dos artefatos pode ser corrigida com ganho na acurácia diagnóstica, sem a necessidade de utilização de técnicas de correção que aumentem a dose de exposição à radiação pelo paciente. A proposta desse estudo foi a criação de um atlas de cintilografia de perfusão miocárdica, que foi obtido a partir de imagens de indíviduos normais, e o desenvolvimento de um algoritmo computacional para a detecção de anormalidades perfusionais miocárdicas, através da comparação estatística dos modelos do atlas com imagens de pacientes. Métodos de corregistro de imagens de mesma modalidade e outras técnicas de processamento de imagens foram estudados e utilizados para a comparação das imagens dos pacientes com o modelo apropriado. Pela análise visual dos modelos, verificou-se a sua validade como imagem representativa de normalidade perfusional. Para avaliação da detecção, a situação dos segmentos miocárdicos (normal ou anormal) indicada pelo algoritmo de detecção foi comparada com a situação apontada no laudo obtido pela concordância de dois especialistas, de modo a se verificar as concordâncias e discordâncias da técnica em relação ao laudo e se obter a significância estatística. Com isso, verificou-se um índice de concordância positiva da técnica em relação ao laudo de aproximadamente 50%, de concordância negativa próxima a 82% e de concordância geral próxima a 68%. O teste exato de Fisher foi aplicado às tabelas de contingência, obtendo-se um valor de p bicaudal inferior a 0,0001, indicando uma probabilidade muito baixa de as concordâncias terem sido obtidas pelo acaso. Melhorias no algoritmo deverão ser implementadas e testes futuros com um padrão-ouro efetivo serão realizados para validação da técnica. / The computer-aided medical imaging analysis allows the quantitative analysis of abnormalities and enhances diagnostic accuracy. This type of analysis is important for nuclear medicine that uses Single Photon Emission Computed Tomography (SPECT), because in the group of three-dimensional data images, subtle patterns of abnormalities often are important clinical findings. However, images can suffer interference from attenuation artifacts of the emission of photons by soft parts of the body, which reduces their diagnostic accuracy. Since there are attenuation parameters computed in a template that allows for comparison with images of a given patient, the artifacts interference can be corrected with a gain in diagnostic accuracy, without the need of using correction techniques that increase the radiation exposure dose of the patient. The purpose of this study was to create an atlas of myocardial perfusion scintigraphy, which was obtained from images of normal individuals and the development of a computational algorithm for detection of myocardial perfusion abnormalities by statistical comparison of atlas templates with images of patients. Methods of image registration of same modality and other image processing techniques were studied and used for comparison of patient images with the appropriate template. By the visual analysis of the templates it was found its validity as a representative image of normal perfusion. For the detection evaluation, the situation of myocardial segments (normal or abnormal) indicated by the detection algorithm was compared with the situation indicated in the medical appraisal report obtained by agreement of two specialists in order to determine the agreement and disagreement of the technique regarding the medical appraisal report and obtaining the statistical significance. Thus, there was a positive agreement index of the technique regarding the medical appraisal report of approximately 50%, a negative agreement index close to 82% and a general agreement index near 68%. The Fisher exact test was applied to the contingency tables, yielding a two-sided p-value less than 0.0001, that indicates a very low probability of the agreements have been obtained by chance. Algorithm improvements should be implemented and further tests with an effective gold-standard will be conducted to validate the technique.