L'obésité est associée à une augmentation de la morbidité et de la mortalité liée à de nombreuses maladies, y compris le diabète de type 2, l'hypertension et des pathologies hépatiques menant à une surcharge lipidique d’origine non alcoolique. Récemment, l’imagerie par résonance magnétique (IRM) est devenue la méthode de choix pour la quantification non invasive de la graisse. Dans cette thèse, les méthodes d'IRM ont été étudiées sur un scanner préclinique de 4.7T in vitro (fantômes MR) et in vivo (souris). Deux algorithmes de quantifications de la graisse -la méthode de Dixon et l’algorithme IDEAL- ont été considérés. Les performances de l'algorithme IDEAL ont été analysées en fonction de propriétés des tissus (T2*, fraction de graisse et modèle spectral de la graisse), de paramètres d'acquisition IRM (temps d’écho, nombre d'échos) et de paramètres expérimentaux (SNR et carte de champ). Sur les fantômes, l'approche standard single-T2* IDEAL a montré certaines limites qui pourraient être surmontées en optimisant le nombre d'échos. Une nouvelle méthode, pour déterminer les valeurs de vérité terrain pour T2* de l'eau et pour T2* de la graisse, a été proposée. Pour les mesures in vivo, différentes analyses ont été effectuées en utilisant l'algorithme IDEAL sur le foie et les muscles. L'analyse statistique sur les mesures de ROI a montré que le choix optimal du nombre d'échos est égal à trois pour la quantification de la graisse et six ou plus pour la quantification du T2*. Les valeurs de la fraction de graisse, calculées avec l'algorithme IDEAL, étaient statistiquement comparables aux valeurs obtenues avec la méthode de Dixon. Enfin, un procédé pour générer des signaux de référence mimant les systèmes eau-graisse (Fat Virtual Phantom MRI), sans l'aide d'objets physiques, a été proposé. Ces fantômes virtuels, qui présentent des caractéristiques de bruit réalistes, représentent une alternative intéressante aux fantômes physiques pour fournir un signal de référence dans les mesures IRM. / Obesity is associated with increased morbidity and mortality linked to many diseases, including type 2 diabetes, hypertension and disease nonalcoholic fatty liver. Recently, 1H magnetic resonance imaging (MRI) has emerged as the method of choice for non-invasive fat quantification. In this thesis, MRI methodologies were investigated for in vitro (MR phantoms) and in vivo (mice) measurements on a 4.7T preclinical scanner. Two algorithms of fat quantifications – the Dixon’s method and IDEAL algorithm – were considered. The performances of the IDEAL algorithm were analyzed as a function of tissue properties (T2*, fat fraction and fat spectral model), MRI acquisition parameters (echo times, number of echoes) and experimental parameters (SNR and field map). In phantoms, the standard approach of single-T2* IDEAL showed some limitations that could be overcome by optimizing the number of echoes. A novel method to determine the ground truth values of T2* of water and T2* of fat was here proposed. For in vivo measurements, different analyses were performed using the IDEAL algorithm in liver and muscle. Statistical analysis on ROI measurements showed that the optimal choice of the number of echoes was equal to three for fat quantification and six or more for T2* quantification. The fat fraction values, calculated with IDEAL algorithm, were statistically similar to the values obtained with Dixon’s method. Finally, a method for generating reference signals mimicking fat-water systems (Fat Virtual Phantom MRI), without using physical objects, was proposed. These virtual phantoms, which display realistic noise characteristics, represent an attractive alternative to physical phantoms for providing a reference signal in MRI measurements.

Yes / Introduction : Rheumatoid arthritis (RA) affects 1% of the population and is principally associated with joint inflammation. It is suggested however that muscle involvement may be one of the earliest clinical features of RA. It is therefore important that techniques exist to accurately assess muscle health in those with RA to enable successful treatment. This study assesses the inter-rater and intra-rater repeatability of Diffusion Tensor MRI (DTI), 2-Point Dixon fat fraction, and T2 relaxation of the thigh muscle in patients with RA using manual regions of interest (ROI). Methods: Nineteen patients (10/19 males; mean age 59; range 18-85) diagnosed with RA had an MRI scan of their hamstrings and quadriceps muscles to obtain fat fraction (FF), mean diffusivity (MD), fractional anisotropy (FA), and T2 quantitative measurements. Two raters (R#1 & R#2) (initials removed for review) independently contoured ROIs for each patient. R#1 repeated the ROI for the same 19 patients after a 6-month hiatus to assess intra-rater repeatability. Inter-rater and intra-rater repeatability for the ROI measurements were compared using Inter Class Correlation (ICC) and Bland-Altman plots. Results: There was excellent agreement for both inter-rater and intra-rater repeatability. ICC results ranged from 0.900-0.998 (P<0.001), and intra-rater ICC results ranged from 0.977-0.999 (P<0.001). Bland-Altman plots also showed excellent agreement. Conclusions: ICC measurements and Bland-Altman plots showed excellent repeatability and agreement with no statistically significant differences when assessing the inter-rater and intra-rater repeatability of FF, MD, FA, and T2 relaxation of the thigh muscle using manual regions of interest in patients with RA. Implications for practice: Manual ROI drawing does not introduce significant errors obtaining FF, MD, FA, and T2 MRI measurements in an RA population. / This research is funded by the NIHR infrastructure at Leeds.

The aim of this exploratory study was to evaluate the influence of hepatic steatosis on the detection rate of metastases in gadoxetic acid-enhanced liver magnetic resonance imaging (MRI). A total of 50 patients who underwent gadoxetic acid-enhanced MRI (unenhanced T1w in- and opposed-phase, T2w fat sat, unenhanced 3D-T1w fat sat and 3-phase dynamic contrast-enhanced (uDP), 3D-T1w fat sat hepatobiliary phase (HP)) were retrospectively included. Two blinded observers (O1/O2) independently assessed the images to determine the detection rate in uDP and HP. The hepatic signal fat fraction (HSFF) was determined as the relative signal intensity reduction in liver parenchyma from in- to opposed-phase images. A total of 451 liver metastases were detected (O1/O2, n = 447/411). O1/O2 detected 10.9%/9.3% of lesions exclusively in uDP and 20.2%/15.5% exclusively in HP. Lesions detected exclusively in uDP were significantly associated with a larger HSFF (area under curve (AUC) of receiver operating characteristic (ROC) analysis, 0.93; p < 0.001; cutoff, 41.5%). The exclusively HP-positive lesions were significantly associated with a smaller diameter (ROC-AUC, 0.82; p < 0.001; cutoff, 5 mm) and a smaller HSFF (ROC-AUC, 0.61; p < 0.001; cutoff, 13.3%). Gadoxetic acid imaging has the advantage of detecting small occult metastatic liver lesions in the HP. However, using non-optimized standard fat-saturated 3D-T1w protocols, severe steatosis (HSFF > 30%) is a potential pitfall for the detection of metastases in HP.

Current methods for estimation of proton density fat fraction (PDFF) of the liver using magnitude magnetic resonance (MR) imaging face the challenge of correctly estimating it when fat is the dominant molecule, i.e. PDFF is more than 50%. Therefore, the accuracy of the methods is limited to half-range operation. We introduce a method based on neural networks for regression capable of estimating over the full range of fat fractions. We built a neural network based on the angles and distances between the data in the discrete MR signal (ADALIFE), using these as features associated to different PDFFs and as input for the network. Tests were performed assessing ADALIFE against dual echo, triple echo, and especially Multi-interference, a state-of-the-art method to estimate PDFFs, with simulated signals at various signal-to-noise (SNR) values. Results were compared in order to verify repeatability and agreement using regression analysis, Bland-Altman and REC curves. Results for Multi-interference were similar to its in-vivo literature, showing the relevance of a simulation. ADALIFE was able to correctly estimate fat fractions up to 100%, breaking the current paradigm for full-range estimation using only off-line post processing. Within half-range, our method outperformed Multi-interference in repeatability and agreement, with narrower limits of agreement and lower expected error at any SNR. / Os métodos atuais para estimação de gordura hepática por densidade de prótons (PDFF) utilizando imagem de magnitude de ressonância magnética (RM) enfrentam o desafio de estimar corretamente quando a gordura é a molécula dominante, ou seja, PDFF é maior que 50%. Assim, a acurácia desses métodos é limitada a meio intervalo de operação. Apresentamos aqui um método baseado em redes neurais para regressão capaz de estimar pelo intervalo completo de frações de gordura. Construímos uma rede neural baseada nos ângulos e distâncias entre os dados do sinal discreto da imagem de RM (ADALIFE), usando esses atributos associados a diferentes valores de PDFF, com sinais simulados considerando diferentes relações sinal-ruído (SNR). Resultados foram comparados para verificar a repetibilidade e concordância através de análise de regressão, Bland- Altman e curvas de característica de erro de regressão (REC). Resultados para o método Multi-interferência (estado-da-arte) foram similares aos relatados in vivo pela literatura, ressaltando a relevância das simulações. ADALIFE foi capaz de estimar corretamente frações de gordura até 100%, quebrando o paradigma para intervalo completo de operação utilizando apenas processamento posterior à aquisição de imagens ou sinais. Considerando meio intervalo, nosso método superou o estado-da-arte em termos de repetibilidade e concordância, com limites mais estreitos e menor erro esperado em qualquer SNR.

Current methods for estimation of proton density fat fraction (PDFF) of the liver using magnitude magnetic resonance (MR) imaging face the challenge of correctly estimating it when fat is the dominant molecule, i.e. PDFF is more than 50%. Therefore, the accuracy of the methods is limited to half-range operation. We introduce a method based on neural networks for regression capable of estimating over the full range of fat fractions. We built a neural network based on the angles and distances between the data in the discrete MR signal (ADALIFE), using these as features associated to different PDFFs and as input for the network. Tests were performed assessing ADALIFE against dual echo, triple echo, and especially Multi-interference, a state-of-the-art method to estimate PDFFs, with simulated signals at various signal-to-noise (SNR) values. Results were compared in order to verify repeatability and agreement using regression analysis, Bland-Altman and REC curves. Results for Multi-interference were similar to its in-vivo literature, showing the relevance of a simulation. ADALIFE was able to correctly estimate fat fractions up to 100%, breaking the current paradigm for full-range estimation using only off-line post processing. Within half-range, our method outperformed Multi-interference in repeatability and agreement, with narrower limits of agreement and lower expected error at any SNR. / Os métodos atuais para estimação de gordura hepática por densidade de prótons (PDFF) utilizando imagem de magnitude de ressonância magnética (RM) enfrentam o desafio de estimar corretamente quando a gordura é a molécula dominante, ou seja, PDFF é maior que 50%. Assim, a acurácia desses métodos é limitada a meio intervalo de operação. Apresentamos aqui um método baseado em redes neurais para regressão capaz de estimar pelo intervalo completo de frações de gordura. Construímos uma rede neural baseada nos ângulos e distâncias entre os dados do sinal discreto da imagem de RM (ADALIFE), usando esses atributos associados a diferentes valores de PDFF, com sinais simulados considerando diferentes relações sinal-ruído (SNR). Resultados foram comparados para verificar a repetibilidade e concordância através de análise de regressão, Bland- Altman e curvas de característica de erro de regressão (REC). Resultados para o método Multi-interferência (estado-da-arte) foram similares aos relatados in vivo pela literatura, ressaltando a relevância das simulações. ADALIFE foi capaz de estimar corretamente frações de gordura até 100%, quebrando o paradigma para intervalo completo de operação utilizando apenas processamento posterior à aquisição de imagens ou sinais. Considerando meio intervalo, nosso método superou o estado-da-arte em termos de repetibilidade e concordância, com limites mais estreitos e menor erro esperado em qualquer SNR.

Pour le diagnostic et la stratification de la fibrose hépatique, la rigidité du foie est un biomarqueur quantitatif estimé par des méthodes d'élastographie. L'élastographie par ondes de cisaillement (« shear wave », SW) utilise des ultrasons médicaux non invasifs pour évaluer les propriétés mécaniques du foie sur la base des propriétés de propagation des ondes de cisaillement. La vitesse des ondes de cisaillement (« shear wave speed », SWS) et l'atténuation des ondes de cisaillement (« shear wave attenuation », SWA) peuvent fournir une estimation de la viscoélasticité des tissus. Les tissus biologiques sont intrinsèquement viscoélastiques et un modèle mathématique complexe est généralement nécessaire pour calculer la viscoélasticité en imagerie SW. Le calcul précis de l'atténuation est essentiel, en particulier pour une estimation précise du module de perte et de la viscosité. Des études récentes ont tenté d'augmenter la précision de l'estimation du SWA, mais elles présentent encore certaines limites. Comme premier objectif de cette thèse, une méthode de décalage de fréquence revisitée a été développée pour améliorer les estimations fournies par la méthode originale de décalage en fréquence [Bernard et al 2017]. Dans la nouvelle méthode, l'hypothèse d'un paramètre de forme décrivant les caractéristiques spectrales des ondes de cisaillement, et assumé initialement constant pour tous les emplacements latéraux, a été abandonnée permettant un meilleur ajustement de la fonction gamma du spectre d'amplitude. En second lieu, un algorithme de consensus d'échantillons aléatoires adaptatifs (« adaptive random sample consensus », A-RANSAC) a été mis en œuvre pour estimer la pente du paramètre de taux variable de la distribution gamma afin d’améliorer la précision de la méthode. Pour valider ces changements algorithmiques, la méthode proposée a été comparée à trois méthodes récentes permettant d’estimer également l’atténuation des ondes de cisaillements (méthodes de décalage en fréquence, de décalage en fréquence en deux points et une méthode ayant comme acronyme anglophone AMUSE) à l'aide de données de simulations ou fantômes numériques. Également, des fantômes de gels homogènes in vitro et des données in vivo acquises sur le foie de canards ont été traités. Comme deuxième objectif, cette thèse porte également sur le diagnostic précoce de la stéatose hépatique non alcoolique (NAFLD) qui est nécessaire pour prévenir sa progression et réduire la mortalité globale. À cet effet, la méthode de décalage en fréquence revisitée a été testée sur des foies humains in vivo. La performance diagnostique de la nouvelle méthode a été étudiée sur des foies humains sains et atteints de la maladie du foie gras non alcoolique. Pour minimiser les sources de variabilité, une méthode d'analyse automatisée faisant la moyenne des mesures prises sous plusieurs angles a été mise au point. Les résultats de cette méthode ont été comparés à la fraction de graisse à densité de protons obtenue de l'imagerie par résonance magnétique (« magnetic resonance imaging proton density fat fraction », MRI-PDFF) et à la biopsie du foie. En outre, l’imagerie SWA a été utilisée pour classer la stéatose et des seuils de décision ont été établis pour la dichotomisation des différents grades de stéatose. Finalement, le dernier objectif de la thèse consiste en une étude de reproductibilité de six paramètres basés sur la technologie SW (vitesse, atténuation, dispersion, module de Young, viscosité et module de cisaillement). Cette étude a été réalisée chez des volontaires sains et des patients atteints de NAFLD à partir de données acquises lors de deux visites distinctes. En conclusion, une méthode robuste de calcul du SWA du foie a été développée et validée pour fournir une méthode de diagnostic de la NAFLD. / For diagnosis and staging of liver fibrosis, liver stiffness is a quantitative biomarker estimated by elastography methods. Ultrasound shear wave (SW) elastography utilizes noninvasive medical ultrasound to assess the mechanical properties of the liver based on the monitoring of the SW propagation. SW speed (SWS) and SW attenuation (SWA) can provide an estimation of tissue viscoelasticity. Biological tissues are inherently viscoelastic in nature and a complex mathematical model is usually required to compute viscoelasticity in SW imaging. Accurate computation of attenuation is critical, especially for accurate loss modulus and viscosity estimation. Recent studies have made attempts to increase the precision of SWA estimation, but they still face some limitations. As a first objective of this thesis, a revisited frequency-shift method was developed to improve the estimates provided by the original implementation of the frequency-shift method [Bernard et al 2017]. In the new method, the assumption of a constant shape parameter of the gamma function describing the SW magnitude spectrum has been dropped for all lateral locations, allowing a better gamma fitting. Secondly, an adaptive random sample consensus algorithm (A-RANSAC) was implemented to estimate the slope of the varying rate parameter of the gamma distribution to improve the accuracy of the method. For the validation of these algorithmic changes, the proposed method was compared with three recent methods proposed to estimate SWA (frequency-shift, two-point frequency-shift and AMUSE methods) using simulation data or numerical phantoms. In addition, in vitro homogenous gel phantoms and in vivo animal (duck) liver data were processed. As a second objective, this thesis also aimed at improving the early diagnosis of nonalcoholic fatty liver disease (NAFLD), which is necessary to prevent its progression and decrease the overall mortality. For this purpose, the revisited frequency-shift method was tested on in vivo human livers. The new method's diagnosis performance was investigated with healthy and NAFLD human livers. To minimize sources of variability, an automated analysis method averaging measurements from several angles has been developed. The results of this method were compared to the magnetic resonance imaging proton density fat fraction (MRI-PDFF) and to liver biopsy. SWA imaging was used for grading steatosis and cut-off decision thresholds were established for dichotomization of different steatosis grades. As a third objective, this thesis is proposing a reproducibility study of six SW-based parameters (speed, attenuation, dispersion, Young’s modulus, viscosity and shear modulus). The assessment was performed in healthy volunteers and NAFLD patients using data acquired at two separate visits. In conclusion, a robust method for computing the liver’s SWA was developed and validated to provide a diagnostic method for NAFLD.