Le cancer de la prostate est le cancer le plus fréquent chez les hommes dans la plupart des pays développés. C'est le cancer le plus fréquent chez les hommes en France (73.609 cas en 2014) et en Colombie (9564 cas en 2014). En outre, c'est la troisième cause de décès par cancer chez les hommes dans les deux pays (9,3 % en France et 7,1 % en Colombie en 2014). L'une des techniques de traitement est la radiothérapie externe, qui consiste à délivrer un rayonnement ionisant à une cible clinique, à savoir la prostate et les vésicules séminales. En raison des variations anatomiques au cours du traitement, qui consiste en environ 40 fractions de rayonnement délivrant une dose totale allant de 70 à 80Gy, des marges de sécurité sont définies autour de la cible tumorale lors de la planification du traitement. Ceci entraîne des portions d'organes sains voisins de la prostate - la vessie et le rectum - à être inclus dans le volume cible, pouvant conduire à des événements indésirables affectant les fonctions urinaires (hématurie et cystite, entre autres) ou rectale (saignement rectal, incontinence fécale, Etc.). La vessie présente les plus grandes variations de forme entre fractions de traitement, provoquées par des changements continus de volume. Ces variations de forme introduisent des incertitudes géométriques qui rendent difficile l'évaluation de la dose réellement délivrée à la vessie pendant le traitement. Ces incertitudes limitent la possibilité de modéliser une relation dose-volume pour la toxicité génito-urinaire tardive (GU). Le projet QUANTEC (Quantitative Analysis of Normal Tissue Effects in the Clinic) a déclaré que la relation dose-réponse pour la toxicité gastro-intestinale tardive (GI) était loin d'être établie. Les variables dosimétriques obtenues à partir de la tomodensitométrie de planification peuvent être faiblement représentative de la dose effectivement administrée. En conséquence, il est crucial de quantifier les incertitudes produites par les variations inter-fraction de la vessie afin de déterminer les facteurs dosimétriques qui affectent les complications GU tardives. Le but de cette thèse était donc de caractériser et de prédire les incertitudes produites par les variations géométriques de la vessie entre les fractions de traitement, en utilisant uniquement la tomodensitométrie de planification comme information d'entrée. En pratique clinique, une seule tomodensitométrie est disponible au moment de la planification du traitement pour un patient typique, alors que des images supplémentaires peuvent être acquises en cours de traitement. Dans cette thèse une approche population a été utilisée pour obtenir suffisamment de données pour apprendre les directions les plus importantes du mouvement et de la déformation de la vessie en utilisant l'analyse en composante principales (ACP). Comme dans les travaux de référence, ces directions ont ensuite été utilisées pour développer des modèles basés population pour prédire et quantifier les incertitudes géométriques de la vessie. Cependant, nous avons utilisé une analyse longitudinale afin de caractériser correctement la variance du patient et les modes spécifiques du patient à partir de la population. Nous avons proposé d'utiliser un modèle à effets mixtes (ME) et une ACP hiérarchique pour séparer la variabilité intra et inter-patients afin de contrôler les effets de cohorte confondus. Finalement, nous avons présenté des modèles sur l'APC comme un outil pour quantifier des incertitudes de la dose produit par le mouvement et déformation de la vessie entre fractions. / Prostate cancer is the most common cancer amongst the male population in most developed countries. It is the most common cancer amongst the male population in France (73.609 cases in 2014) and in Colombia (9564 cases in 2014). It is also the third most common cause of cancer deaths in males in both countries (9.3% and 7.1% in France and in Colombia in 2014, respectively). One of the standard treatment methods is external radiotherapy, which involves delivering ionizing radiation to a clinical target, namely the prostate and seminal vesicles. Due to the uncertain location of organs during treatment, which involves around forty (40) radiation fractions delivering a total dose ranging from 70 to 80Gy, safety margins are defined around the tumor target upon treatment planning. This leads to portions of healthy organs neighboring the prostate or organs at risk — the bladder and rectum — to be included in the target volume, potentially resulting in adverse events affecting patients’ urinary (hematuria and cystitis, among others) or rectal (rectal bleeding, fecal incontinence, etc.) functions. The bladder is notorious for presenting the largest inter-fraction shape variations during treatment, caused by continuous changes in volume. These variations in shape introduce geometric uncertainties that render assessment of the actual dose delivered to the bladder during treatment difficult, thereby leading to dose uncertainties that limit the possibility of modeling dose-volume response for late genitourinary (GU) toxicity. The Quantitative Analysis of Normal Tissue Effects in the Clinic (QUANTEC) project has stated that a similar dose-response to that of late gastrointestinal (GI) toxicity is far from being established. The dosimetric variables obtained from the planning CT prove to be very poor surrogates for the real delivered dose. As a result, it appears crucial to quantify uncertainties produced by inter-fraction bladder variations in order to determine dosimetric factors that affect late GU complications. The aim of this thesis was thus to characterize and predict uncertainties produced by geometric variations of the bladder between fractions, using solely the planning CT as input information. In clinical practice, a single CT scan is only available for a typical patient during the treatment planning while on-treatment CTs/CBCTs are seldom available. In this thesis, we thereby used a population approach to obtain enough data to learn the most important directions of bladder motion and deformation using principal components analysis (PCA). As in groundwork, these directions were then used to develop population-based models in order to predict and quantify geometrical uncertainties of the bladder. However, we use a longitudinal analysis in order to properly characterize both patient-specific variance and modes from the population. We proposed to use mixed-effects (ME) models and hierarchical PCA to separate intra and inter-patient variability to control confounding cohort effects. . Subsequently, we presented PCA models as a tool to quantify dose uncertainties produced by bladder motion and deformation between fractions.

This study examines the relationship between gully erosion in channels, sidewalls, and interfluves, and precipitation parameters (duration, total accumulation, average intensity, and maximum intensity) annually and seasonally to determine seasonal drivers for precipitation-related erosion. Ordinary Least Square regression models of erosion using precipitation and antecedent precipitation at weekly lags of up to twelve weeks were developed for three erosion variables for each of three geomorphic areas: channels, interfluves, and sidewalls (nine models in total). Erosion was most pronounced in winter months, followed by spring, indicating the influence of high-intensity precipitation from frontal systems and repeated freeze-thaw cycles in winter; erosion in summer was driven by high-intensity precipitation from convectional storms. Annually, duration was the most important driver for erosion, however, during winter and summer months, precipitation intensity was dominant. Seasonal models retained average and maximum precipitation as drivers for erosion in winter months (dominated by frontal systems), and retained maximum precipitation intensity as a driver for erosion in summer months (dominated by convectional storms). In channels, precipitation duration was the dominant driver for erosion due to runoff-related erosion, while in sidewalls and interfluves intensity parameters were equally important as duration, likely related to rain splash erosion. These results show that the character of precipitation, which varies seasonally, is an important driver for gully erosion and that studies of precipitation-driven erosion should consider partitioning data by season to identify these drivers.

This study examines gully erosion in northeast Tennessee hillslopes in the Southern Appalachian Valley and Ridge physiographic province, where a thick sequence of red clay Ultisols (Acrisol, according to the World Reference Base for Soil) overlies dolomite and limestone bedrock. The role of freeze-thaw processes in gully erosion was examined weekly from 6/3/2012 to 9/17/2014 using a network of n = 78 erosion pins in three geomorphic areas: channels, interfluves, and sidewalls. Freeze-thaw days were identified using meteorological data collected on site. When freeze-thaw days occurred, erosion and deposition increased and gully conditions were more dynamic. When daily temperature did not plunge below freezing, more stable gully conditions persisted. Ordinary Least Square regression models of erosion pin length using freeze-thaw events explained significant portions of variability in channels (R² = 0.113, p < 0.01), interfluves (R² = 0.141, p < 0.01), and sidewalls (R² = 0.263, p < 0.01). Repeat analysis on only the winter-spring months minimally improved the sidewall model (R² = 0.272, p < 0.01). Erosion in interfluves exhibited a lagged effect, and was best correlated to freeze-thaw events during the prior period while erosion in channels and sidewalls was related to freeze-thaw events in the current week. Of the three geomorphic areas studied, sidewall erosion was best modeled by freeze-thaw events which contribute to widening of gullies through mobilization of sediment and mass wasting. This research demonstrates that freeze-thaw processes are a significant contributor to erosion in gully channels, interfluves, and especially sidewalls, and therefore temperature variability should be considered in erosion studies in similar climates.

The contourlet transform is an emerging multiscale multidirection image processing technique. It effectively represents smooth curvature details typical of natural images, overcoming a major drawback of the 2-D wavelet transform. To further exploit its potential, in this research, a statistical model, the contourlet contextual hidden Markov model (C-CHMM), has been developed to characterize contourlet images. A systematic mutual information based context construction procedure has been developed to form an appropriate context for the model. With this contourlet image model, a multiscale segmentation method has also been established for the application to texture images. The segmentation method combines a model comparison approach with a multiscale fusion and a multi-neighbor combination process. It also features a neighborhood selection scheme based on a smoothed context map, for both the model estimation and the neighbor combination. The effectiveness of the image model has been verified through a series of denoising and segmentation experiments. As demonstrated with the denoising performance, this new model for contourlet images is more promising than the state of the art, the contourlet hidden Markov tree (C-HMT) model. The other model being compared with in this work is the wavelet contextual hidden Markov model (W-CHMM). Through the denoising experiments, the presented C-CHMM shows better robustness against noise than the W-CHMM. Moreover, the new model demonstrates its superiority to the wavelet model in the segmentation performance. Through the segmentation experiments, the value of the systematic context construction procedure has been proven. The C-CHMM based segmentation method has also been validated. In comparison with the state of the art methods for the same type, the presented technique shows improved accuracy in segmenting texture patterns of diversified nature. This success in segmentation has further manifested the potential of the newly developed contourlet image model.

No description available.

No description available.

Organ transplantation is one of the most important and effective solutions to save end-stage patients, who have one or more critical organ failures. However, the inadequate organs for transplantation to meet the demands has been the major issue. Even worse, the lack of accurate non-invasive assessment methods wastes 20% of donor organs every year. Currently, the most frequently used organ assessment methods are visual inspections and biopsy. Yet both methods are subjective: the assessment accuracy depends on the evaluator's experience. Moreover, repeating biopsies will potentially damage the organs. To reduce the waste of donor organs, online non-invasive and quantitative organ assessment methods are in great needs. Organ viability assessment is a challenging issue due to four reasons: 1) there are no universally accepted guidelines or procedures for surgeons to quantitatively assess the organ viability; 2) there is no easy-deployed and non-invasive biological in situ data to correlate with organ viability; 3) the organs viability is difficult to model because of heterogeneity among organs; 4) both visual inspection and biopsy can be applied only at present time, and how to forecast the viability of similar-but-non-identical organs at a future time is still in shadow. Motivated by the challenges, the overall objective of this dissertation is to develop online non-invasive and quantitative assessment methods to predict and forecast the organ viability. As a result, four data-driven modeling research tasks are investigated to achieve the overall objective: 1) Quantitative and qualitative models are used to jointly predict the number of dead cells and the liver viability based on features extracted from biopsy images. This method can quantitatively assess the organ viability, which could be used to validate the biopsy results from pathologists to increase the evaluation accuracy. 2) A multitask learning logistic regression model is applied to assess liver viability by using principal component analysis to extract infrared image features to quantify the correlation between liver viability and spatial infrared imaging data. This non-invasive online assessment method can evaluate the organ viability without physical contact to reduce the risk of damaging the organs. 3) A spatial-temporal smooth variable selection method is conducted to improve the liver viability prediction accuracy by considering both spatial and temporal effects from the infrared images without feature engineering. In addition, it provides medical interpretation based on variable selection to highlight the most significant regions on the liver resulting in viability loss. 4) A multitask general path model is implemented to forecast the heterogeneous kidney viability based on limited historical data by learning the viability loss paths of each kidney during preservation. The generality of this method is validated by tissue deformation forecasting in needle biopsy process to potentially improve the biopsy accuracy. In summary, the proposed data-driven methods can predict and forecast the organ viability without damaging the organ. As a result, the increased utilization rate of donor organs will benefit more end-stage patients by dramatically extending their life spans. / Doctor of Philosophy / Organ transplantation is the ultimate solution to save end-stage patients with one or more organ failures. However, the inadequate organs for transplantation to meet the demands has been the major issue. Even worse, the lack of accurate and non-invasive viability assessment methods wastes 20% of donor organs every year. Currently, the most frequently used organ assessment methods are visual inspections and biopsy. Yet both methods are subjective: the assessment accuracy depends on the personal experience of evaluator. Moreover, repeating biopsies will potentially damage the organs. As a result, online non-invasive and quantitative organ assessment methods are in great needs. It is extremely important because such methods will increase the organ utilization rate by saving more discarded organs with transplantation potential. The overall objective of this dissertation is to advance the knowledge on modeling organ viability by developing online non-invasive and quantitative methods to predict and forecast the viability of heterogeneous organs in transplantation. After an introduction in Chapter 1, four research tasks are investigated. In Chapter 2, quantitative and qualitative models jointly predicting porcine liver viability are proposed based on features from biopsy images to validate the biopsy results. In Chapter 3, a multi-task learning logistic regression model is proposed to assess the cross-liver viability by correlating liver viability with spatial infrared data validated by porcine livers. In Chapter 4, a spatial-temporal smooth variable selection is proposed to predict liver viability by considering both spatial and temporal correlations in modeling without feature engineering, which is also validated by porcine livers. In addition, the variable selection results provide medical interpretations by capturing the significant regions on the liver in predicting viability. In Chapter 5, a multitask general path model is proposed to forecast kidney viability validated by porcine kidney. This forecasting method is generalized to apply to needle biopsy tissue deformation case study with the objective to improve the needle insertion accuracy. Finally, I summarize the research contribution and discuss future research directions in Chapter 6. The proposed data-driven methods can predict and forecast organ viability without damaging the organ. As a result, the increased utilization rate of donor organs will benefit more patients by dramatically extending their life spans and bringing them back to normal daily activities.

Cette thèse porte sur la segmentation d'images OCT cutanées. Cette modalité d'imagerie permet de visualiser les structures superficielles avec une profondeur de l'ordre du millimètre. En dermatologie, elle permet d'explorer l'épiderme et sa jonction avec le derme. Cependant, les images OCT sont sévèrement affectées par le bruit speckle. Ce phénomène conjugué à la complexité inhérente aux structures de la peau rend l'interprétation des images difficile même pour des experts. L'approche classique consiste à filtrer le speckle avant de faire des traitements de segmentation. A l'inverse, dans cette thèse nous exploitons exclusivement le bruit comme information pour segmenter. Notre approche repose sur la modélisation statistique du speckle. La segmentation se fait par classification des paramètres de ce modèle probabiliste. Ainsi, - On montre que le speckle ne suit pas une loi Rayleigh, comme cela est établi analytiquement. - On ajuste plusieurs lois de probabilité à l'amplitude OCT; et on montre que celle-ci est distribuée selon la loi Gamma généralisée. - On établit que les paramètres de cette loi discriminent statistiquement les couches de la peau. - On conçoit une méthode de segmentation par classification des paramètres locaux du speckle. Les nombreuses expérimentations faites sur plusieurs corpus d'images in-vivo confirment la validité de notre approche. / This thesis deals with the segmentation of skin OCT images. This modality provides the means to visualize superficial structures down to a millimeter depth. In dermatology, it is used to examine the epidermis and its junction with the dermis. However, OCT images are severely affected by the speckle noise. This random phenomenon added to the complexity of human skin structures makes the visual interpretation of images very difficult. Classical image processing techniques filter this noise prior to any segmentation step. In this thesis, we rely exclusively on the properties of the speckle to perform segmentation. Our approach is based on the statistical modeling of the speckle. Images are segmented by classifying parameters of the statistical model. Therefore, - We show that speckle does not follow the Rayleigh distribution, as developed analytically in the literature. - We fit various probability laws to model OCT signal amplitude ; we show that Generalized Gamma has the best goodness of fit. - We establish that statistical parameters of this distribution discriminate skin layers with good variability. - We develop a segmentation method based on the classification of local statistical parameters. The various experimental results with a number of in-vivo images reported in the thesis confirm the validity of our approach

The study of social network dynamics has become an increasingly important component of many disciplines in the social sciences. In the past decade, statistical models and methods have been proposed which permit researchers to draw statistical inference on these dynamics. This thesis builds on one such family of models, the stochastic actor oriented model (SAOM) proposed by Snijders [2001]. Goodness of fit for SAOMs is an area that is only just beginning to be filled in with appropriate methods. This thesis proposes a Mahalanobis distance based, Monte Carlo goodness of fit test that can depend on arbitrary features of the observed network data and covariates. As remediating poor fit can be a difficult process, a modified model distance (MMD) estimator is devised that can help researchers to choose among a set of model elaborations. In practice, panel data is typically used to draw SAOM-based inference. This thesis also proposes a score-type test for time heterogeneity between the waves in the panel that is computationally cheap and fits into a convenient, forward model selecting workflow. Next, this thesis proposes a rigorous method for aggregating so-called relational event data (e.g. emails and phone calls) by extending the SAOM family to a family of hidden Markov models that suppose a latent social network is driving the observed relational events. Finally, this thesis proposes a measurement model for SAOMs inspired by error-in-variables (EiV) models employed in an array of disciplines. Like the relational event aggregation model, the measurement model is a hidden Markov model extension to the SAOM family. These models allow the researcher to specify the form of the mesurement error and buffer against potential attenuating biases and other problems that can arise if the errors are ignored.

Hydraulic fracturing is used to extract oil and natural gas from low permeability formations. The potential of fluids migrating from depth through adjacent wellbores and through the production wellbore was investigated using statistical modeling and predic-tive classifiers. The probability of a hydraulic fracturing well becoming hydraulically connected to an adjacent well in the Marcellus shale of New York was determined to be between 0.00% and 3.45% at the time of the study. This means that the chance of an in-duced fracture from hydraulic fracturing intersecting an existing well is highly dependent on the area of increased permeability caused by fracturing. The chance of intersecting an existing well does not mean that fluid will flow upwards; for upward migration to occur, a pathway must exist and a pressure gradient is required to drive flow, with the exception of gas flow caused by buoyancy. Predictive classifiers were employed on a dataset of wells in Alberta Canada to identify well characteristics most associated to fluid migration along the production well. The models, specifically a random forest, were able to identify pathways better than random guessing with 78% of wells in the data set identified cor-rectly. Magnetic resonance imaging (MRI) was used to visualize and quantify contami-nant transport in a soil column using a full body scanner. T1 quantification was used to determine the concentration of a contaminant surrogate in the form of Magnevist, an MRI contrast agent. Imaging showed a strong impact from density driven convection when the density difference between the two fluids was small (0.3%). MRI also identified a buildup of contrast agent concentration at the interface between a low permeability ground silica and higher permeability AFS 50-70 testing sand when density driven con-vection was eliminated.