Submitted by Johnny Rodrigues (johnnyrodrigues@ufcg.edu.br) on 2018-05-01T14:25:43Z No. of bitstreams: 1 ADALBERTO GOMES TEIXEIRA JÚNIOR - DISSERTAÇÃO PPGCC 2015..pdf: 2611686 bytes, checksum: 6b9c4a2efc3946611ad0263328434bd1 (MD5) / Made available in DSpace on 2018-05-01T14:25:43Z (GMT). No. of bitstreams: 1 ADALBERTO GOMES TEIXEIRA JÚNIOR - DISSERTAÇÃO PPGCC 2015..pdf: 2611686 bytes, checksum: 6b9c4a2efc3946611ad0263328434bd1 (MD5) Previous issue date: 2015-07 / Capes / O som gerado por motores em funcionamento contém informações sobre seu estado e condições, tornando-se uma fonte importante para a avaliação de seu funcionamento sem a necessidade de intervenção no equipamento. A análise do estado do equipamento muitas vezes é realizada por diagnóstico humano, a partir da experiência vivenciada no ambiente ruidoso de operação. Como o funcionamento dos motores é regido por um processo periódico, o sinal de áudio gerado segue um padrão bem definido, possibilitando, assim, a avaliação de seu estado de funcionamento por meio desse sinal. Dentro deste contexto, a pesquisa ora descrita trata da modelagem do sinal acústico gerado por motores em usinas termoelétricas, aplicando técnicas de processamento digital de sinais e inteligência artificial, com o intuito de auxiliar o diagnóstico de falhas, minimizando a presença humana no ambiente de uma sala de motores. A técnica utilizada baseia-se no estudo do funcionamento dos equipamentos e dos sinais acústicos por eles gerados por esses, para a extração de características representativas do sinal, em diferentes domínios, combinadas a métodos de aprendizagem de máquinas para a construção de um multiclassificador, responsável pela avaliação do estado de funcionamento desses motores. Para a avaliação da eficácia do método proposto, foram utilizados sinais extraídos de motores da Usina Termoelétrica Borborema Energética S.A., no âmbito do projeto REPARAI (REPair over AiR using Artificial Intelligence, código ANEEL PD6471-0002/2012). Ao final do estudo, o método proposto demonstrou acurácia próxima a 100%. A abordagem proposta caracterizou-se, portanto, como eficiente para o diagnóstico de falhas, principalmente por não ser um método invasivo, não exigindo, portanto, o contato direto do avaliador humano com o motor em funcionamento. / The sound generated by an engine during operation contains information about its conditions, becoming an important source of information to evaluate its status without requiring intervention in equipment. The fault diagnosis of the engine usually is performed by a human, based on his experience in a noisy environment. As the operation of the engine is a periodic procedure, the generated signal follows a well-defined pattern, allowing the evaluation of its operating conditions. On this context, this research deals with modeling the acoustic signal generated by engines in power plants, using techniques from digital signal processing and artificial intelligence, with the purpose of assisting the fault diagnosis, minimizing the human presence at the engine room. The technique applied is based on the study of engines operation and the acoustic signal generated by them, extracting signal representative characteristics in different domains, combined with machine learning methods, to build a multiclassifier to evaluate the engines status. Signals extracted from engines of Borborema Energética S.A. power plant, during the REPARAI Project (REPair over AiR using Artificial Intelligence), ANEEL PD-6471-0002/2012, were used in the experiments. In this research, the method proposed has demonstrated an accuracy rate of nearly 100%. The approach has proved itself to be efficient to fault diagnosis, mainly by not being an invasive method and not requiring human direct contact with the engine.

Cette thèse est une collaboration entre le LAboratoire des Sciences et Matériaux pour l’Électronique et d’Automatique (LASMEA) de Clermont-Ferrand et le Laboratoire Vision et Ingénierie des Contenus (LVIC) du CEA LIST à Saclay. La première moitié de la thèse a été accomplie au sein de l’équipe ComSee (1) du LASMEA et la deuxième au LVIC. L’objectif de ces travaux est de concevoir un système de vidéo-assistance temps réel pour la détection d’évènements dans des scènes possiblement denses.La vidéosurveillance intelligente de scènes denses telles que des foules est particulièrement difficile, principalement à cause de leur complexité et de la grande quantité de données à traiter simultanément. Le but de cette thèse consiste à élaborer une méthode de détection d’évènements rares dans de telles scènes, observées depuis une caméra fixe. La méthode en question s’appuie sur l’analyse automatique de mouvement et ne nécessite aucune information à priori. Les mouvements nominaux sont déterminés grâce à un apprentissage statistique non supervisé. Les plus fréquemment observés sont considérés comme des évènements normaux. Une phase de classification permet ensuite de détecter les mouvements déviant trop du modèle statistique, pour les considérer comme anormaux. Cette approche est particulièrement adaptée aux lieux de déplacements structurés, tels que des scènes de couloirs ou de carrefours routiers. Aucune étape de calibration, de segmentation de l’image, de détection d’objets ou de suivi n’est nécessaire. Contrairement aux analyses de trajectoires d’objets suivis, le coût calculatoire de notre méthode est invariante au nombre de cibles présentes en même temps et fonctionne en temps réel. Notre système s’appuie sur une classification locale du mouvement de la scène, sans calibration préalable. Dans un premier temps, une caractérisation du mouvement est réalisée, soit par des méthodes classiques de flot optique, soit par des descripteurs spatio-temporels. Ainsi, nous proposons un nouveau descripteur spatio-temporel fondé sur la recherche d’une relation linéaire entre les gradients spatiaux et les gradients temporels en des zones où le mouvement est supposé uniforme. Tout comme les algorithmes de flot optique, ce descripteur s’appuie sur la contrainte d’illumination constante.Cependant en prenant en compte un voisinage temporel plus important, il permet une caractérisation du mouvement plus lisse et plus robuste au bruit. De plus, sa faible complexité calculatoire est bien adaptée aux applications temps réel. Nous proposons ensuite d’étudier différentes méthodes de classification : La première, statique, dans un traitement image par image, s’appuie sur une estimation bayésienne de la caractérisation du mouvement au travers d’une approche basée sur les fenêtres de Parzen. Cette nouvelle méthode est une variante parcimonieuse des fenêtres de Parzen. Nous montrons que cette approche est algorithmiquement efficace pour approximer de manière compacte et précise les densités de probabilité. La seconde méthode, basée sur les réseaux bayésiens, permet de modéliser la dynamique du mouvement. Au lieu de considérer ce dernier image par image, des séquences de mouvements sont analysées au travers de chaînes de Markov Cachées. Ajouté à cela, une autre contribution de ce manuscrit est de prendre en compte la modélisation du voisinage d’un bloc afin d’ajouter une cohérence spatiale à la propagation du mouvement. Ceci est réalisé par le biais de couplages de chaînes de Markov cachées.Ces différentes approches statistiques ont été évaluées sur des données synthétiques ainsi qu’en situations réelles, aussi bien pour la surveillance du trafic routier que pour la surveillance de foule.Cette phase d’évaluation permet de donner des premières conclusions encourageantes quant à la faisabilité de la vidéosurveillance intelligente d’espaces possiblement denses. / The automatic analysis of crowded areas in video sequences is particularly difficult because ofthe large amount of information to be processed simultaneously and the complexity of the scenes. We propose in this thesis a method for detecting abnormal events in possibly dense scenes observed from a static camera. The approach is based on the automatic classification of motion requiring no prior information. Motion patterns are encoded in an unsupervised learning framework in order to generate a statistical model of frequently observed (aka. normal) events. Then at the detection stage, motion patterns that deviate from the model are classified as unexpected events. The method is particularly adapted to scenes with structured movement with directional flow of objects or people such as corridors, roads, intersections. No camera calibration is needed, nor image segmentation, object detection and tracking. In contrast to approaches that rely on trajectory analysis of tracked objects, our method is independent of the number of targets and runs in real-time. Our system relies on a local classification of global scene movement. The local analysis is done on each blocks of a regular grid. We first introduce a new spatio-temporal local descriptor to characterize the movement efficiently. Assuming a locally uniform motion of space-time blocks of the image, our approach consists in determining whether there is a linear relationship between spatial gradients and temporal gradients. This spatio-temporal descriptor holds the Illumination constancy constraint like optical flow techniques, but it allows taking into account the spatial neighborhood and a temporal window by giving a smooth characterization of the motion, which makes it more robust to noise. In addition, its low computational complexity is suitable for real-time applications. Secondly, we present two different classification frameworks : The first approach is a static (frame by frame) classification approach based on a Bayesian characterization of the motion by using an approximation of the Parzen windowing method or Kernel Density Estimation (KDE) to model the probability density function of motion patterns.This new method is the sparse variant of the KDE (SKDE). We show that the SKDE is a very efficient algorithm giving compact representations and good approximations of the density functions. The second approach, based on Bayesian Networks, models the dynamics of the movement. Instead of considering motion patterns in each block independently, temporal sequences of motion patterns are learned by using Hidden Markov Models (HMM). The second proposed improvement consists in modeling the movement in one block by taking into account the observed motion in adjacent blocks. This is performed by the coupled HMM method. Evaluations were conducted to highlight the classification performance of the proposed methods,on both synthetic data and very challenging real video sequences captured by video surveillance cameras.These evaluations allow us to give first conclusions concerning automatic analyses of possibly crowded area.

Thesis is aimed at the possibility of utilization of extreme learning machines and echo state networks for time series forecasting with possibility of utilizing GPU acceleration. Such predictions are part of nearly everyone’s daily lives through utilization in weather forecasting, prediction of regular and stock market, power consumption predictions and many more. Thesis is meant to familiarize reader firstly with theoretical basis of extreme learning machines and echo state networks, taking advantage of randomly generating majority of neural networks parameters and avoiding iterative processes. Secondly thesis demonstrates use of programing tools, such as ND4J and CUDA toolkit, to create very own programs. Finally, prediction capability and convenience of GPU acceleration is tested.

The correct classification of a logs assortment is crucial for the economic output within a fully mechanized timber harvest. This task is especially for unexperienced but also for professional machine operators mentally demanding. This paper presents a method towards an assistance system for machine operators for an automated log quality and assortment estimation. Therefore, machine vision methods for object detection are combined with machine learning approaches for estimating the logs weight based on a Convolutional Neural Network (CNN). Based on the dimensions oft he object ´log, a first categorisation into a specific assortment is done. By comparing the theoretical weight of a healthy log of such dimensions to the real weight estimated by the CNN-based crane scale, quality reducing properties such as beetle infestation or red rod can be detected. In such cases, the assistance system displays a visual warning to the operator to check the loaded log.

No description available.

Indiana University-Purdue University Indianapolis (IUPUI) / The objective of this thesis is to compare automated methods for performing analysis of flow cytometry data. Flow cytometry is an important and efficient tool for analyzing the characteristics of cells. It is used in several fields, including immunology, pathology, marine biology, and molecular biology. Flow cytometry measures light scatter from cells and fluorescent emission from dyes which are attached to cells. There are two main tasks that must be performed. The first is the adjustment of measured fluorescence from the cells to correct for the overlap of the spectra of the fluorescent markers used to characterize a cell’s chemical characteristics. The second is to use the amount of markers present in each cell to identify its phenotype. Several methods are compared to perform these tasks. The Unconstrained Least Squares, Orthogonal Subspace Projection, Fully Constrained Least Squares and Fully Constrained One Norm methods are used to perform compensation and compared. The fully constrained least squares method of compensation gives the overall best results in terms of accuracy and running time. Spectral Clustering, Gaussian Mixture Modeling, Naive Bayes classification, Support Vector Machine and Expectation Maximization using a gaussian mixture model are used to classify cells based on the amounts of dyes present in each cell. The generative models created by the Naive Bayes and Gaussian mixture modeling methods performed classification of cells most accurately. These supervised methods may be the most useful when online classification is necessary, such as in cell sorting applications of flow cytometers. Unsupervised methods may be used to completely replace manual analysis when no training data is given. Expectation Maximization combined with a cluster merging post-processing step gives the best results of the unsupervised methods considered.

The global transmission of Severe Acute Respiratory Syndrome Coronavirus-2 (SARS-CoV-2) has resulted in over 677 million infections and 6.88 million tragic deaths worldwide as of March 10th, 2023. During the pandemic, the ability to effectively combat SARS-CoV-2 had been hindered by the lack of rapid, reliable, and cost-effective testing platforms for readily screening patients, discerning incubation stages, and accounting for variants. The limited knowledge of the viral pathogenesis further hindered rapid diagnosis and long-term clinical management of this complex disease. While effective in the short term, measures such as social distancing and lockdowns have resulted in devastating economic loss, in addition to material and psychological hardships. Therefore, successfully reopening society during a pandemic depends on frequent, reliable testing, which can result in the timely isolation of highly infectious cases before they spread or become contagious. Viral loads, and consequently an individual's infectiousness, change throughout the progression of the illness. These dynamics necessitate frequent testing to identify when an infected individual can safely interact with non-infected individuals. Thus, scalable, accurate, and rapid serial testing is a cornerstone of an effective pandemic response, a prerequisite for safely reopening society, and invaluable for early containment of epidemics. Given the significant challenges posed by the pandemic, the power of artificial intelligence (AI) can be harnessed to create new diagnostic methods and be used in conjunction with serial tests. With increasing utilization of at-home lateral flow immunoassay (LFIA) tests, the National Institutes of Health (NIH) and Centers for Disease Control and Prevention (CDC) have consistently raised concerns about a potential underreporting of actual SARS-CoV-2-positive cases. When AI is paired with serial tests, it could instantly notify, automatically quantify, aid in real-time contact tracing, and assist in isolating infected individuals. Moreover, the computer vision-assisted methodology can help objectively diagnose conditions, especially in cases where subjective LFIA tests are employed. Recent advances in the interdisciplinary scientific fields of machine learning and biomedical engineering support a unique opportunity to design AI-based strategies for pandemic preparation and response. Deep learning algorithms are transforming the interpretation and analysis of image data when used in conjunction with biomedical imaging modalities such as MRI, Xray, CT scans, confocal microscopes, etc. These advances have enabled researchers to carry out real-time viral infection diagnostics that were previously thought to be impossible. The objective of this thesis is to use SARS-CoV-2 as a model virus and investigate the potential of applying multi-class instance segmentation deep learning and other machine learning strategies to build pandemic preparedness for rapid, in-depth, and longitudinal diagnostic platforms. This thesis encompasses three research tasks: 1) computer vision-assisted rapid serial testing, 2) infected cell phenotyping, and 3) diagnosing the long-term consequences of infection (i.e., long-term COVID). The objective of Task 1 is to leverage the power of AI, in conjunction with smartphones, to rapidly and simultaneously diagnose COVID-19 infections for millions of people across the globe. AI not only makes it possible for rapid and simultaneous screenings of millions but can also aid in the identification and contact tracing of individuals who may be carriers of the virus. The technology could be used, for example, in university settings to manage the entry of students into university buildings, ensuring that only students who test negative for the virus are allowed within campus premises, while students who test positive are placed in quarantine until they recover. The technology could also be used in settings where strict adherence to COVID-19 prevention protocols is compromised, for example, in an Emergency Room. This technology could also help with CDC’s concern on growing incidences of underreporting positive COVID-19 cases with growing utilization of at-home LFIA tests. AI can address issues that arise from relying solely on the visual interpretation of LFIA tests to make accurate diagnoses. One problem is that LFIA test results may be subjective or ambiguous, especially when the test line of the LFIA displays faint color, indicating a low analyte abundance. Therefore, reaching a decisive conclusion regarding the patient's diagnosis becomes challenging. Additionally, the inclusion of a secondary source for verifying the test results could potentially increase the test's cost, as it may require the purchase of complementary electronic gadgets. To address these issues, our innovation would be accurately calibrated with appropriate sensitivity markers, ensuring increased accuracy of the diagnostic test and rapid acquisition of test results from the simultaneous classification of millions of LFIA tests as either positive or negative. Furthermore, the designed network architecture can be utilized to detect other LFIA-based tests, such as early pregnancy detection, HIV LFIA detection, and LFIA-based detection of other viruses. Such minute advances in machine learning and artificial intelligence can be leveraged on many different scales and at various levels to revolutionize the health sector. The motivating purpose of Task 2 is to design a highly accurate instance segmentation network architecture not only for the analysis of SARS-CoV-2 infected cells but also one that yields the highest possible segmentation accuracy for all applications in biomedical sciences. For example, the designed network architecture can be utilized to analyze macrophages, stem cells, and other types of cells. Task 3 focuses on conducting studies that were previously considered computationally impossible. The invention will assist medical researchers and dentists in automatically calculating alveolar crest height (ACH) in teeth using over 500 dental Xrays. This will help determine if patients diagnosed with COVID-19 by a positive PCR test exhibited more alveolar bone loss and had greater bone loss in the two years preceding their COVID-positive test when compared to a control group without a positive COVID-19 test. The contraction of periodontal disease results in higher levels of transmembrane serine protease 2 (TMPRSS2) within the buccal cavity, which is instrumental in enabling the entry of SARS-CoV-2. Gum inflammation, a symptom of periodontal disease, can lead to alterations in the ACH of teeth within the oral mucosa. Through this innovation, we can calculate ACHs of various teeth and, therefore, determine the correlation between ACH and the risk of contracting SARS-CoV-2 infection. Without the invention, extensive manpower and time would be required to make such calculations and gather data for further research into the effects of SARS-CoV-2 infection, as well as other related biological phenomena within the human body. Furthermore, the novel network framework can be modified and used to calculate dental caries and other periodontal diseases of interest.

Backdoor (Trojan) attacks are a major threat to the security of deep neural network (DNN) models. They are created by an attacker who adds a certain pattern to a portion of given training dataset, causing the DNN model to misclassify any inputs that contain the pattern. These infected classifiers are called Trojan models and the added pattern is referred to as the trigger. In image domain, a trigger can be a patch of pixel values added to the images and in text domain, it can be a set of words. In this thesis, we propose Trojan-Miner (T-Miner), a defense scheme against such backdoor attacks on text classification deep learning models. The goal of T-Miner is to detect whether a given classifier is a Trojan model or not. To create T-Miner , our approach is based on a sequence-to-sequence text generation model. T-Miner uses feedback from the suspicious (test) classifier to perturb input sentences such that their resulting class label is changed. These perturbations can be different for each of the inputs. T-Miner thus extracts the perturbations to determine whether they include any backdoor trigger and correspondingly flag the suspicious classifier as a Trojan model. We evaluate T-Miner on three text classification datasets: Yelp Restaurant Reviews, Twitter Hate Speech, and Rotten Tomatoes Movie Reviews. To illustrate the effectiveness of T-Miner, we evaluate it on attack models over text classifiers. Hence, we build a set of clean classifiers with no trigger in their training datasets and also using several trigger phrases, we create a set of Trojan models. Then, we compute how many of these models are correctly marked by T-Miner. We show that our system is able to detect trojan and clean models with 97% overall accuracy over 400 classifiers. Finally, we discuss the robustness of T-Miner in the case that the attacker knows T-Miner framework and wants to use this knowledge to weaken T-Miner performance. To this end, we propose four different scenarios for the attacker and report the performance of T-Miner under these new attack methods. / M.S. / Backdoor (Trojan) attacks are a major threat to the security of predictive models that make use of deep neural networks. The idea behind these attacks is as follows: an attacker adds a certain pattern to a portion of given training dataset and in the next step, trains a predictive model over this dataset. As a result, the predictive model misclassifies any inputs that contain the pattern. In image domain this pattern that is called trigger, can be a patch of pixel values added to the images and in text domain, it can be a set of words. In this thesis, we propose Trojan-Miner (T-Miner), a defense scheme against such backdoor attacks on text classification deep learning models. The goal of T-Miner is to detect whether a given classifier is a Trojan model or not. T-Miner is based on a sequence-to-sequence text generation model that is connected to the given predictive model and determine if the predictive model is being backdoor attacked. When T-Miner is connected to the predictive model, it generates a set of words, called perturbations, and analyses these perturbations to determine whether they include any backdoor trigger. Hence if any part of the trigger is present in the perturbations, the predictive model is flagged as a Trojan model. We evaluate T-Miner on three text classification datasets: Yelp Restaurant Reviews, Twitter Hate Speech, and Rotten Tomatoes Movie Reviews. To illustrate the effectiveness of T-Miner, we evaluate it on attack models over text classifiers. Hence, we build a set of clean classifiers with no trigger in their training datasets and also using several trigger phrases, we create a set of Trojan models. Then, we compute how many of these models are correctly marked by T-Miner. We show that our system is able to detect Trojan models with 97% overall accuracy over 400 predictive models.

La maladie d’Alzheimer (AD) est la principale cause de démence neurodégénérative, caractérisée à 95% par des formes tardives (LOAD) qui présentent des troubles mnésiques et progressent lentement. Cependant, environ 5% des patients atteints d’AD présentent une forme précoce de la maladie (EOAD) débutant avant 65 ans. Bien que le substratum lésionnel soit identique à la LOAD, l’EOAD est caractérisée par une plus grande sévérité des dépôts de plaques amyloïdes, des enchevêtrements neurofibrillaires et de l’atrophie cérébrale. De plus, l’EOAD est plus hétérogène que la LOAD, car même si la majorité des troubles sont mnésiques il existe une proportion importante de formes atypiques affectées par des troubles du langage, visuospatiaux ou exécutifs. Bien que de nombreuses études en imagerie TEP 18F-FDG aient permis de caractériser métaboliquement l’EOAD par rapport à la LOAD ou à un groupe de contrôles sains, très peu différentiaient pas les formes typiques (mnésiques) des formes atypiques. Dans ce travail de thèse, nous avons examiné les données d’imagerie TEP 18F-FDG, complémentées par l’IRM structurelle, afin d’améliorer la caractérisation et la compréhension des formes typiques et atypiques d’EOAD. Suite à un premier travail d’harmonisation des reconstructions TEP 18F-FDG entre deux machines GE et Siemens ayant toutes deux servies à l’acquisition des données patients, notre second objectif a été d’étudier à l’inclusion sur le cerveau entier les patterns hypométaboliques caractéristiques des différentes formes d’EOAD et leurs corrélations potentielles avec la performance neuropsychologique. Cette étude a montré que chaque forme clinique d’EOAD était caractérisée par des patterns hypométaboliques spécifiques fortement corrélés aux symptômes cliniques et aux scores neuropsychologiques du domaine cognitif associé. Par la suite, nous nous sommes intéressés à la progression sur 3 ans de l’hypométabolisme sur la surface corticale en fonction des formes typiques ou atypiques d’EOAD. Bien que des patterns similaires d’évolution de l’hypométabolisme entre les formes typiques et atypiques aient été observés au niveau du cortex pariétal, seules les formes atypiques ont présenté une réduction du métabolisme bilatérale plus importante au niveau du cortex orbito-frontal latéral associée à des déclins cognitifs plus sévères. Temporellement, les résultats suggèrent que l’hypométabolisme chez les formes typiques progresserait selon un axe antérieur-vers-postérieur en cohérence avec les stades de Braak et Braak, alors que l’hypométabolisme chez les formes atypiques progresserait selon un axe postérieur-vers-antérieur. Pris ensemble, ces résultats confortent l’hypothèse d’une distribution différente de la pathologie tau en termes de charge et d’évolution temporelle entre ces deux formes d’EOAD. Notre dernier objectif a été de déterminer les capacités discriminatives des données TEP 18F-FDG, seules ou combinées aux données de l’IRM structurelle, afin de classifier de manière automatique et supervisée des patients atteints d’EOAD en forme typique ou atypique. Nous avons mis en application des algorithmes de machine learning combinés à des méthodes de validation croisée afin d’évaluer les influences de diverses composantes sur les performances de classification. Des précisions équilibrées maximales égales à 80,8% en imagerie monomodale TEP 18F-FDG et 92,4% en imagerie multimodale TEP 18F-FDG/IRM T1 ont été obtenues, validant ainsi la TEP 18F-FDG comme un biomarqueur sensible de l’EOAD et soulignant l’apport incontestable de la multimodalité. En conclusion, nos travaux ont permis une meilleure caractérisation et compréhension des formes cliniques d’EOAD, ouvrant la voie à un management personnalisé du patient et des traitements plus efficaces pour ces formes distinctes. / Alzheimer’s disease (AD) is the most common form of neurodegenerative dementia, characterized at 95% by late-onset forms (LOAD) which present episodic memory impairments and progress slowly. However, 5% of AD patients have an early-onset form (EOAD) of the disease whose onset begins before 65. Although the lesion substratum is similar between EOAD and LOAD, EOAD has more severe neuritic plaque deposits, neurofibrillary tangles and brain atrophy. Moreover, EOAD is more heterogeneous than LOAD, because even if most of the impairments are about episodic memory there is a high proportion of atypical forms impaired in language, visuospatial or executive functions. Although many 18F-FDG PET studies allowed to metabolically characterize EOAD compared to LOAD or healthy controls group, very few differentiated typical from atypical forms. In this thesis, we examined 18F-FDG PET data, complemented by structural MRI, in order to improve characterization and comprehension of typical and atypical forms of EOAD. Following a first harmonization work between 18F-FDG PET reconstructions from both GE and Siemens scanners used for the acquisition of patient data, our second aim was to study at baseline on the whole brain hypometabolic patterns characterizing the clinical forms of EOAD and their correlations with neuropsychological performance. This work showed that each clinical form of EOAD was characterized by specific hypometabolic patterns highly correlated with clinical symptoms and neuropsychological performance of the associated cognitive domain. Then, we focused on the 3-year hypometabolism progression on the cortical surface according typical or atypical forms of EOAD. Although similar patterns of hypometabolism evolution between typical and atypical forms were observed in parietal cortices, atypical only showed a more severe reduction of metabolism in lateral orbitofrontal cortices associated with more severe cognitive declines. Temporally, the results suggest that hypometabolism in typical forms would progress according to an anterior-to-posterior axis coherently with Braak and Braak stages, whereas in atypical forms hypometabolism would progress according a posterior-to-anterior axis. Taken together, results consolidate the hypothesis of a different tau distribution in terms of burden and temporal evolution between both forms of EOAD. Our last goal was to determine the discriminative power of 18F-FDG PET data, alone or combined to structural MRI data, in order to automatically classify in a supervised manner EOAD patients into typical or atypical form. We applied machine learning algorithms combined to cross-validation methods to assess influence of some components on classification performances. Maximum balanced accuracies equal to 80.8% in monomodal 18F-FDG PET and 92.4% in multimodal 18F-FDG PET/T1 MRI were obtained, validating 18F-FDG PET as a sensible biomarker of EOAD and highlighting the incontestable contribution of multimodality. In conclusion, our works allowed a better characterization and comprehension of clinical forms of EOAD, paving the way to personalized patient management and more effective treatments for these distinct clinical forms.

L'imagerie par résonance magnétique fonctionnelle (IRMf) permet de mesurer l'activité cérébrale à travers le flux sanguin apporté aux neurones. Dans cette thèse nous évaluons la capacité de modèles biologiquement plausibles et issus de la vision par ordinateur à représenter le contenu d'une image de façon similaire au cerveau. Les principaux modèles de vision évalués sont les réseaux convolutionnels.Les réseaux de neurones profonds ont connu un progrès bouleversant pendant les dernières années dans divers domaines. Des travaux antérieurs ont identifié des similarités entre le traitement de l'information visuelle à la première et dernière couche entre un réseau de neurones et le cerveau. Nous avons généralisé ces similarités en identifiant des régions cérébrales correspondante à chaque étape du réseau de neurones. Le résultat consiste en une progression des niveaux de complexité représentés dans le cerveau qui correspondent à l'architecture connue des aires visuelles: Plus la couche convolutionnelle est profonde, plus abstraits sont ses calculs et plus haut niveau sera la fonction cérébrale qu'elle sait modéliser au mieux. Entre la détection de contours en V1 et la spécificité à l'objet en cortex inférotemporal, fonctions assez bien comprises, nous montrons pour la première fois que les réseaux de neurones convolutionnels de détection d'objet fournissent un outil pour l'étude de toutes les étapes intermédiaires du traitement visuel effectué par le cerveau.Un résultat préliminaire à celui-ci est aussi inclus dans le manuscrit: L'étude de la réponse cérébrale aux textures visuelles et sa modélisation avec les réseaux convolutionnels de scattering.L'autre aspect global de cette thèse sont modèles de “décodage”: Dans la partie précédente, nous prédisions l'activité cérébrale à partir d'un stimulus (modèles dits d’”encodage”). La prédiction du stimulus à partir de l'activité cérébrale est le méchanisme d'inférence inverse et peut servir comme preuve que cette information est présente dans le signal. Le plus souvent, des modèles linéaires généralisés tels que la régression linéaire ou logistique ou les SVM sont utilisés, donnant ainsi accès à une interprétation des coefficients du modèle en tant que carte cérébrale. Leur interprétation visuelle est cependant difficile car le problème linéaire sous-jacent est soit mal posé et mal conditionné ou bien non adéquatement régularisé, résultant en des cartes non-informatives. En supposant une organisation contigüe en espace et parcimonieuse, nous nous appuyons sur la pénalité convexe d'une somme de variation totale et la norme L1 (TV+L1) pour développer une pénalité regroupant un terme d'activation et un terme de dérivée spatiale. Cette pénalité a la propriété de mettre à zéro la plupart des coefficients tout en permettant une variation libre des coefficients dans une zone d'activation, contrairement à TV+L1 qui impose des zones d’activation plates. Cette méthode améliore l'interprétabilité des cartes obtenues dans un schéma de validation croisée basé sur la précision du modèle prédictif.Dans le contexte des modèles d’encodage et décodage nous tâchons à améliorer les prétraitements des données. Nous étudions le comportement du signal IRMf par rapport à la stimulation ponctuelle : la réponse impulsionnelle hémodynamique. Pour générer des cartes d'activation, au lieu d’un modèle linéaire classique qui impose une réponse impulsionnelle canonique fixe, nous utilisons un modèle bilinéaire à réponse hémodynamique variable spatialement mais fixe à travers les événements de stimulation. Nous proposons un algorithme efficace pour l'estimation et montrons un gain en capacité prédictive sur les analyses menées, en encodage et décodage. / Blood-oxygen-level dependent (BOLD) functional magnetic resonance imaging (fMRI) makes it possible to measure brain activity through blood flow to areas with metabolically active neurons. In this thesis we use these measurements to evaluate the capacity of biologically inspired models of vision coming from computer vision to represent image content in a similar way as the human brain. The main vision models used are convolutional networks.Deep neural networks have made unprecedented progress in many fields in recent years. Even strongholds of biological systems such as scene analysis and object detection have been addressed with enormous success. A body of prior work has been able to establish firm links between the first and last layers of deep convolutional nets and brain regions: The first layer and V1 essentially perform edge detection and the last layer as well as inferotemporal cortex permit a linear read-out of object category. In this work we have generalized this correspondence to all intermediate layers of a convolutional net. We found that each layer of a convnet maps to a stage of processing along the ventral stream, following the hierarchy of biological processing: Along the ventral stream we observe a stage-by-stage increase in complexity. Between edge detection and object detection, for the first time we are given a toolbox to study the intermediate processing steps.A preliminary result to this was obtained by studying the response of the visual areas to presentation of visual textures and analysing it using convolutional scattering networks.The other global aspect of this thesis is “decoding” models: In the preceding part, we predicted brain activity from the stimulus presented (this is called “encoding”). Predicting a stimulus from brain activity is the inverse inference mechanism and can be used as an omnibus test for presence of this information in brain signal. Most often generalized linear models such as linear or logistic regression or SVMs are used for this task, giving access to a coefficient vector the same size as a brain sample, which can thus be visualized as a brain map. However, interpretation of these maps is difficult, because the underlying linear system is either ill-defined and ill-conditioned or non-adequately regularized, resulting in non-informative maps. Supposing a sparse and spatially contiguous organization of coefficient maps, we build on the convex penalty consisting of the sum of total variation (TV) seminorm and L1 norm (“TV+L1”) to develop a penalty grouping an activation term with a spatial derivative. This penalty sets most coefficients to zero but permits free smooth variations in active zones, as opposed to TV+L1 which creates flat active zones. This method improves interpretability of brain maps obtained through cross-validation to determine the best hyperparameter.In the context of encoding and decoding models, we also work on improving data preprocessing in order to obtain the best performance. We study the impulse response of the BOLD signal: the hemodynamic response function. To generate activation maps, instead of using a classical linear model with fixed canonical response function, we use a bilinear model with spatially variable hemodynamic response (but fixed across events). We propose an efficient optimization algorithm and show a gain in predictive capacity for encoding and decoding models on different datasets.