Machine learning is evidenced as a research area with the main purpose of developing computational methods that are capable of learning with their previously acquired experiences. Although a large amount of machine learning techniques has been proposed and successfully applied in real systems, there are still many challenging issues, which need be addressed. In the last years, an increasing interest in techniques based on complex networks (large-scale graphs with nontrivial connection patterns) has been verified. This emergence is explained by the inherent advantages provided by the complex network representation, which is able to capture the spatial, topological and functional relations of the data. In this work, we investigate the new features and possible advantages offered by complex networks in the machine learning domain. In fact, we do show that the network-based approach really brings interesting features for supervised, semisupervised, and unsupervised learning. Specifically, we reformulate a previously proposed particle competition technique for both unsupervised and semisupervised learning using a stochastic nonlinear dynamical system. Moreover, an analytical analysis is supplied, which enables one to predict the behavior of the proposed technique. In addition to that, data reliability issues are explored in semisupervised learning. Such matter has practical importance and is found to be of little investigation in the literature. With the goal of validating these techniques for solving real problems, simulations on broadly accepted databases are conducted. Still in this work, we propose a hybrid supervised classification technique that combines both low and high orders of learning. The low level term can be implemented by any classification technique, while the high level term is realized by the extraction of features of the underlying network constructed from the input data. Thus, the former classifies the test instances by their physical features, while the latter measures the compliance of the test instances with the pattern formation of the data. Our study shows that the proposed technique not only can realize classification according to the semantic meaning of the data, but also is able to improve the performance of traditional classification techniques. Finally, it is expected that this study will contribute, in a relevant manner, to the machine learning area / Aprendizado de máquina figura-se como uma área de pesquisa que visa a desenvolver métodos computacionais capazes de aprender com a experiência. Embora uma grande quantidade de técnicas de aprendizado de máquina foi proposta e aplicada, com sucesso, em sistemas reais, existem ainda inúmeros problemas desafiantes que necessitam ser explorados. Nos últimos anos, um crescente interesse em técnicas baseadas em redes complexas (grafos de larga escala com padrões de conexão não triviais) foi verificado. Essa emergência é explicada pelas inerentes vantagens que a representação em redes complexas traz, sendo capazes de capturar as relações espaciais, topológicas e funcionais dos dados. Nesta tese, serão investigadas as possíveis vantagens oferecidas por redes complexas quando utilizadas no domínio de aprendizado de máquina. De fato, será mostrado que a abordagem por redes realmente proporciona melhorias nos aprendizados supervisionado, semissupervisionado e não supervisionado. Especificamente, será reformulada uma técnica de competição de partículas para o aprendizado não supervisionado e semissupervisionado por meio da utilização de um sistema dinâmico estocástico não linear. Em complemento, uma análise analítica de tal modelo será desenvolvida, permitindo o entendimento evolucional do modelo no tempo. Além disso, a questão de confiabilidade de dados será investigada no aprendizado semissupervisionado. Tal tópico tem importância prática e é pouco estudado na literatura. Com o objetivo de validar essas técnicas em problemas reais, simulações computacionais em bases de dados consagradas pela literatura serão conduzidas. Ainda nesse trabalho, será proposta uma técnica híbrica de classificação supervisionada que combina tanto o aprendizado de baixo como de alto nível. O termo de baixo nível pode ser implementado por qualquer técnica de classificação tradicional, enquanto que o termo de alto nível é realizado pela extração das características de uma rede construída a partir dos dados de entrada. Nesse contexto, aquele classifica as instâncias de teste segundo qualidades físicas, enquanto que esse estima a conformidade da instância de teste com a formação de padrões dos dados. Os estudos aqui desenvolvidos mostram que o método proposto pode melhorar o desempenho de técnicas tradicionais de classificação, além de permitir uma classificação de acordo com o significado semântico dos dados. Enfim, acredita-se que este estudo possa gerar contribuições relevantes para a área de aprendizado de máquina.

本研究的主要目的是檢視倒傳導神經網路是否具有人類在分類學習上所呈現出來的學習效應 — 競爭學習、遮蔽效應與不相關線索的影響。在實驗中，我們採用兩種倒傳導神經網路，來測試激發函數是否會影響倒傳導神經網路的學習。此兩種倒傳導神經網路分別採用sigmoid激發函數與hyperbolic-tangent激發函數。實驗結果顯示，以sigmoid為激發函數與以hyperbolic-tangent為激發函數的倒傳導神經網路都具有這三個學習效應。還有，以sigmoid為激發函數的倒傳導神經網路所呈現出來的學習效應比以hyperbolic-tangent為激發函數的倒傳導神經網路來得顯著。本研究的次要目的在於瞭解有效性（使用性）與敏感度分析的數值是否有對應關係。實驗結果顯示，線索A與線索B的敏感度分析數值差異可以反映出線索A與線索B的有效性差異。然而，敏感度分析數值卻無法準確地顯示線索的有效性數值。 / The main objective of this research is to examine whether back propagation neural networks (BP) have the learning effects found in human category learning — competitive learning, overshadowing and the deleterious of an irrelevant cue. Two kinds of BP, BP with sigmoid activation function and BP with hyperbolic-tangent activation function, are investigated to see if the activation function will make BP behave differently. According to the results of our experiments, these three learning effects are demonstrated both in BP with sigmoid and BP with hyperbolic-tangent, but they seems more significant in BP with sigmoid than in BP with hyperbolic-tangent. The second objective of our research is to see if there is a correspondence between the validity (the utilization) and the value of sensitivity analysis, R. From the results of our experiments, we observe that the difference between values of sensitivity analysis with respect to Cue A and Cue B reflects the difference of the validities between Cue A and Cue B. However, the value of sensitivity analysis does not show exactly what validity a cue is.

Cette thèse est issue d’un accord CIFRE entre le groupe de recherche EnCoV de l’Université Clermont Auvergne et la société Quantel Medical (www.quantel-medical.fr). Quantel Medical est une entreprise spécialisée dans le développement innovant des ultrasons et des produits laser en ophtalmologie. Cette thèse présente un travail de recherche visant à l’application du diagnostic assisté par ordinateur et du traitement des maladies de la rétine avec une utilisation du prototype industriel TrackScan développé par Quantel Medical. Plus précisément, elle contribue au problème du mosaicing précis de l’image de la lampe à fente (SLIM) et du recalage automatique et multimodal en utilisant les images SLIM avec l’angiographie par fluorescence (FA) pour aider à la photo coagulation pan-rétienne naviguée. Nous abordons trois problèmes différents.Le premier problème est lié à l’accumulation des erreurs du recalage en SLIM., il dérive de la mosaïque. Une approche commune pour obtenir la mosaïque consiste à calculer des transformations uniquement entre les images temporellement consécutives dans une séquence, puis à les combiner pour obtenir la transformation entre les vues non consécutives temporellement. Les nombreux algorithmes existants suivent cette approche. Malgré le faible coût de calcul et la simplicité de cette méthode, en raison de sa nature de ‘chaînage’, les erreurs d’alignement s’accumulent, ce qui entraîne une dérive des images dans la mosaïque. Nous proposons donc d’utilise les récents progrès réalisés dans les méthodes d’ajustement de faisceau et de présenter un cadre de réduction de la dérive spécialement conçu pour SLIM. Nous présentons aussi une nouvelle procédure de raffinement local.Deuxièmement, nous abordons le problème induit par divers types d’artefacts communs á l’imagerie SLIM. Ceus-sont liés à la lumière utilisée, qui dégrade considérablement la qualité géométrique et photométrique de la mosaïque. Les solutions existantes permettent de faire face aux blouissements forts qui corrompent entièrement le rendu de la rétine dans l’image tout en laissant de côté la correction des reflets spéculaires semi-transparents et reflets des lentilles. Cela introduit des images fantômes et des pertes d’information. En outre, les méthodes génériques ne produisent pas de résultats satisfaisants dans SLIM. Par conséquent, nous proposons une meilleure alternative en concevant une méthode basée sur une technique rapide en utilisant une seule image pour éliminer les éblouissements et la notion de feux spéculaires semi-transparents en utilisant les indications de mouvement pour la correction intelligente de reflet de lentille.Finalement, nous résolvons le problème du recalage multimodal automatique avec SLIM. Il existe une quantité importante de travaux sur le recalage multimodal de diverses modalités d’image rétinienne. Cependant, la majorité des méthodes existantes nécessitent une détection de points clés dans les deux modalités d’image, ce qui est une tâche très difficile. Dans le cas de SLIM et FA ils ne tiennent pas compte du recalage précis dans la zone maculaire - le repère prioritaire. En outre, personne n’a développé une solution entièrement automatique pour SLIM et FA. Dans cette thèse, nous proposons la première méthode capable de recolu ces deux modalités sans une saisie manuelle, en détectant les repères anatomiques uniquement sur une seule image pour assurer un recalage précis dans la zone maculaire. (...) / This thesis arises from an agreement Convention Industrielle de Formation par la REcherche (CIFRE) between the Endoscopy and Computer Vision (EnCoV) research group at Université Clermont Auvergne and the company Quantel Medical (www.quantel-medical.fr), which specializes in the development of innovative ultrasound and laser products in ophthalmology. It presents a research work directed at the application of computer-aided diagnosis and treatment of retinal diseases with a use of the TrackScan industrial prototype developed at Quantel Medical. More specifically, it contributes to the problem of precise Slit-Lamp Image Mosaicing (SLIM) and automatic multi-modal registration of SLIM with Fluorescein Angiography (FA) to assist navigated pan-retinal photocoagulation. We address three different problems.The first is a problem of accumulated registration errors in SLIM, namely the mosaicing drift.A common approach to image mosaicking is to compute transformations only between temporally consecutive images in a sequence and then to combine them to obtain the transformation between non-temporally consecutive views. Many existing algorithms follow this approach. Despite the low computational cost and the simplicity of such methods, due to its ‘chaining’ nature, alignment errors tend to accumulate, causing images to drift in the mosaic. We propose to use recent advances in key-frame Bundle Adjustment methods and present a drift reduction framework that is specifically designed for SLIM. We also introduce a new local refinement procedure.Secondly, we tackle the problem of various types of light-related imaging artifacts common in SLIM, which significantly degrade the geometric and photometric quality of the mosaic. Existing solutions manage to deal with strong glares which corrupt the retinal content entirely while leaving aside the correction of semi-transparent specular highlights and lens flare. This introduces ghosting and information loss. Moreover, related generic methods do not produce satisfactory results in SLIM. Therefore, we propose a better alternative by designing a method based on a fast single-image technique to remove glares and the notion of the type of semi-transparent specular highlights and motion cues for intelligent correction of lens flare.Finally, we solve the problem of automatic multi-modal registration of FA and SLIM. There exist a number of related works on multi-modal registration of various retinal image modalities. However, the majority of existing methods require a detection of feature points in both image modalities. This is a very difficult task for SLIM and FA. These methods do not account for the accurate registration in macula area - the priority landmark. Moreover, none has developed a fully automatic solution for SLIM and FA. In this thesis, we propose the first method that is able to register these two modalities without manual input by detecting retinal features only on one image and ensures an accurate registration in the macula area.The description of the extensive experiments that were used to demonstrate the effectiveness of each of the proposed methods is also provided. Our results show that (i) using our new local refinement procedure for drift reduction significantly ameliorates the to drift reduction allowing us to achieve an improvement in precision over the current solution employed in the TrackScan; (ii) the proposed methodology for correction of light-related artifacts exhibits a good efficiency, significantly outperforming related works in SLIM; and (iii) despite our solution for multi-modal registration builds on existing methods, with the various specific modifications made, it is fully automatic, effective and improves the baseline registration method currently used on the TrackScan.

Machine learning is evidenced as a research area with the main purpose of developing computational methods that are capable of learning with their previously acquired experiences. Although a large amount of machine learning techniques has been proposed and successfully applied in real systems, there are still many challenging issues, which need be addressed. In the last years, an increasing interest in techniques based on complex networks (large-scale graphs with nontrivial connection patterns) has been verified. This emergence is explained by the inherent advantages provided by the complex network representation, which is able to capture the spatial, topological and functional relations of the data. In this work, we investigate the new features and possible advantages offered by complex networks in the machine learning domain. In fact, we do show that the network-based approach really brings interesting features for supervised, semisupervised, and unsupervised learning. Specifically, we reformulate a previously proposed particle competition technique for both unsupervised and semisupervised learning using a stochastic nonlinear dynamical system. Moreover, an analytical analysis is supplied, which enables one to predict the behavior of the proposed technique. In addition to that, data reliability issues are explored in semisupervised learning. Such matter has practical importance and is found to be of little investigation in the literature. With the goal of validating these techniques for solving real problems, simulations on broadly accepted databases are conducted. Still in this work, we propose a hybrid supervised classification technique that combines both low and high orders of learning. The low level term can be implemented by any classification technique, while the high level term is realized by the extraction of features of the underlying network constructed from the input data. Thus, the former classifies the test instances by their physical features, while the latter measures the compliance of the test instances with the pattern formation of the data. Our study shows that the proposed technique not only can realize classification according to the semantic meaning of the data, but also is able to improve the performance of traditional classification techniques. Finally, it is expected that this study will contribute, in a relevant manner, to the machine learning area / Aprendizado de máquina figura-se como uma área de pesquisa que visa a desenvolver métodos computacionais capazes de aprender com a experiência. Embora uma grande quantidade de técnicas de aprendizado de máquina foi proposta e aplicada, com sucesso, em sistemas reais, existem ainda inúmeros problemas desafiantes que necessitam ser explorados. Nos últimos anos, um crescente interesse em técnicas baseadas em redes complexas (grafos de larga escala com padrões de conexão não triviais) foi verificado. Essa emergência é explicada pelas inerentes vantagens que a representação em redes complexas traz, sendo capazes de capturar as relações espaciais, topológicas e funcionais dos dados. Nesta tese, serão investigadas as possíveis vantagens oferecidas por redes complexas quando utilizadas no domínio de aprendizado de máquina. De fato, será mostrado que a abordagem por redes realmente proporciona melhorias nos aprendizados supervisionado, semissupervisionado e não supervisionado. Especificamente, será reformulada uma técnica de competição de partículas para o aprendizado não supervisionado e semissupervisionado por meio da utilização de um sistema dinâmico estocástico não linear. Em complemento, uma análise analítica de tal modelo será desenvolvida, permitindo o entendimento evolucional do modelo no tempo. Além disso, a questão de confiabilidade de dados será investigada no aprendizado semissupervisionado. Tal tópico tem importância prática e é pouco estudado na literatura. Com o objetivo de validar essas técnicas em problemas reais, simulações computacionais em bases de dados consagradas pela literatura serão conduzidas. Ainda nesse trabalho, será proposta uma técnica híbrica de classificação supervisionada que combina tanto o aprendizado de baixo como de alto nível. O termo de baixo nível pode ser implementado por qualquer técnica de classificação tradicional, enquanto que o termo de alto nível é realizado pela extração das características de uma rede construída a partir dos dados de entrada. Nesse contexto, aquele classifica as instâncias de teste segundo qualidades físicas, enquanto que esse estima a conformidade da instância de teste com a formação de padrões dos dados. Os estudos aqui desenvolvidos mostram que o método proposto pode melhorar o desempenho de técnicas tradicionais de classificação, além de permitir uma classificação de acordo com o significado semântico dos dados. Enfim, acredita-se que este estudo possa gerar contribuições relevantes para a área de aprendizado de máquina.