Les métastases sont une cause majeure de décès dans le cas du cancer. En effet, ces tumeurs secondaires peuvent se développer dans divers organes, distants de la tumeur primaire, et surviennent à des temps différents au cours de la croissance de la tumeur primaire. De nombreuses techniques d’imagerie biomédicale peuvent être utilisées pour les détecter. Parmi celles-ci, l’Imagerie par Résonance Magnétique (IRM) a l’avantage de ne pas utiliser de rayonnements ionisants et permet de forts contrastes entre des tissus mous différents. Toutefois il est encore nécessaire de développer de nouvelles techniques IRM pour mieux caractériser les tumeurs, obtenir des données quantitatives, et réduire drastiquement les durées d’examen. Parmi les caractéristiques biophysiques mesurables par IRM, le temps de relaxation T1 semble être un bio-marqueur de l’efficacité d’une thérapie anti-cancéreuse. Cependant, sa mesure est généralement trop chronophage pour être utilisée en imagerie préclinique sur des cohortes d’animaux, ou pour une utilisation en routine clinique. Cette thèse a porté sur le développement d’une séquence de quantification T1 fiable et rapide. Elle a été appliquée pour détecter et caractériser des métastases chez la souris avec comme pré-requis de générer des images avec une résolution spatiale élevée, d’être insensible aux mouvements respiratoires et de permettre des mesures reproductibles du T1. La séquence Magnetization Prepared 2 Rapid Acquisition Gradient Echoes (MP2RAGE) à encodage cartésien a donc été choisie pour son fort potentiel de quantification T1, sa robustesse aux hétérogénéités de champ magnétique, et pour obtenir rapidement des cartes paramétriques en 3D. L’influence des paramètres de la séquence a d’abord été évaluée par simulations. Puis la séquence a été modifiée pour être compatible avec une méthode d’accélération appelée acquisition comprimée. Cette méthode a alors été utilisée soit pour réduire le temps d’acquisition des cartes T1, soit pour en améliorer la résolution spatiale. Cette nouvelle séquence MP2RAGE a alors été utilisée à 7T pour détecter et caractériser des métastases cérébrales disséminées dans le cerveau de souris. Pour détecter des métastases hépatiques, l’encodage cartésien initial s’est avéré trop sensible aux mouvements respiratoires. Il a donc été remplacé par un encodage radial, nécessitant une adaptation du schéma de reconstruction des cartes T1. Ainsi, des cartes T1 3D de l’abdomen entier de souris ont été obtenues en 9 minutes. Un suivi longitudinal de métastases hépatiques a montré des hétérogénéités de T1 inter- et intra-métastatiques. Pour une accélération supplémentaire, la séquence a été développée avec un encodage multi-coupe 2D, permettant ainsi d’utiliser les nombreux temps-morts présents dans le chronogramme. Des optimisations des paramètres de la séquence ont permis d’obtenir 6 cartes T1 en 9 s in vivo sur le cerveau et l’abdomen de souris. De plus, une étude préliminaire a montré qu’elle permettait de réaliser de la thermométrie. Une première perspective de ces travaux consiste à transférer cette séquence sur un aimant de recherche clinique. Une autre perspective serait de développer une séquence multi-coupe 3D radiale, accélérée par acquisition comprimée, applicable sur le petit-animal comme chez l’humain. Celapermettrait d’allier efficacité de la séquence, forte résolution spatiale et robustesse aux mouvements pour un large éventail d’applications. / Metastases are a leading cause of death in the case of cancer. Indeed, these secondary tumors can develop in various organs, distant from the primary tumor, and occur at different times during the growth of the primary tumor. Many biomedical imaging techniques can be used to detect them. Among these, Magnetic Resonance Imaging (MRI) has the advantage of not using ionizing radiation and allows strong contrasts between different soft tissues. However, it remains necessary to develop new MRI techniques to better characterize tumors, obtain quantitative data and to drastically reduce exam times. Among the biophysical characteristics measurable by MRI, the T1 relaxation time seems to be a biomarker of the efficiency of an anti-cancer therapy. However, its measurement is generally too time consuming to be used in preclinical imaging on cohorts of animals, or for routine clinical use. This thesis therefore had the challenge of developing a reliable and rapid T1 quantification sequence. It has been applied to detect anc characterize metastases in mice with the prerequisites of generating images with high spatial resolution, being insensitive to respiratory movements and allowing reproducible T1 measurements. The Magnetization Prepared 2 Rapid Acquisition Gradient Echoes sequence (MP2RAGE) with Cartesian encoding was therefore chosen for its high T1 quantitation potential, its robustness to magnetic field heterogeneities, and its ability to quickly obtain 3D parametric maps. First, simulations were performed to evaluate the influence of sequence parameters. Then the sequence was modified to be compatible with an acceleration method called Compressed Sensing. This method was then used either to reduce the acquisition time of the T1 maps, or to improve the spatial resolution. This new MP2RAGE sequence was then applied at 7T to detect and characterize disseminated brain metastases in the mouse. The initial Cartesian encoding proved to be too sensitive to respiratory movements to detect liver metastases. It was therefore replaced by a radial encoding, which required an adaptation of the reconstruction scheme of the T1 maps. Thus, 3D T1 maps of the entire abdomen of mice were obtained in 9 minutes. Longitudinal follow-up of hepatic metastases showed inter and intra-metastatic T1 heterogeneities. For an additional acceleration, the sequence was developed with a 2D multi-slice encoding, thus allowing to use the many dead times present in the chronogram. Optimizations of the parameters of the sequence made it possible to obtain 6 T1 maps in 9 s in vivo on the brain and the abdomen of mice. In addition, a preliminary study showed a possible application for thermometry. A first perspective of this work is to transfer this sequence to a clinical research MRI system. Another perspective would be to develop a 3D radial multi-slice sequence, accelerated by Compressed Sensing, applicable to small animals as in humans. This would combine sequence efficiency, high spatial resolution and robustness to movements for a wide range of applications.

In this study, three different MR pulse sequences, IR-FSE, MP2RAGE, and firstly proposed MP3RAGE, were applied to obtain high-resolution 3D T1 mapping of whole brain at 1.5 Tesla. Among these three sequences, MP2RAGE uses fast gradient echo as readout module. Signals of two different inversion times are acquired at once and can be used to calculate T1 relaxation time according to Bloch equation. However, the magnetization was also influenced by the excitation efficiency of inversion adiabatic pulse, which was usually estimated by numerical simulation and taken as a constant over the field of view in the literature. However, this might not be true in practice. Therefore, a newly modified pulse sequence, MP3RAGE, was proposed to acquire data of three distinct inversion times without increasing scanning time. As a result, the spatial distribution of T1 and inversion efficiency can be assessed by solving nonlinear least square problem. In addition, the IR-FSE sequence with six inversion times was also applied in every experiment to provide T1 value for reference. Results showed that the T1 estimation obtained by MP2RAGE is close to, but slightly lower than that by IR-FSE, which is in agreement with those reported in literatures. In addition, the 3D high-resolution maps of T1 and efficiency were successfully estimated with the use of MP3RAGE. Spatial smoothing on inversion efficiency helps reducing the sensitivity to noise in the nonlinear approach, leading to T1 values closer to those by IR-FSE.

京都大学 / 0048 / 新制・課程博士 / 博士(医学) / 甲第20257号 / 医博第4216号 / 新制||医||1020(附属図書館) / 京都大学大学院医学研究科医学専攻 / (主査)教授 宮本 享, 教授 村井 俊哉, 教授 高橋 淳 / 学位規則第4条第1項該当 / Doctor of Medical Science / Kyoto University / DFAM

Les travaux réalisés dans cette thèse se situent à l'interface des domaines de l'acquisition en imagerie par résonance magnétique (IRM) et du traitement d'image pour l'analyse automatique des structures cérébrales. La mesure de modifications structurelles telles que l'atrophie corticale nécessite l'application d'algorithmes de traitement d'image. Ceux-ci doivent compenser les artefacts en IRM tels que l'inhomogénéité du signal ou les effets de volume partiel (VP) pour permettre la segmentation des tissus cérébraux puis l'estimation d'épaisseur corticale. Nous proposons une nouvelle modélisation de VP proche de la physique de l'acquisition baptisée modèle bi-exponentiel qui vient concurrencer le traditionnel modèle linéaire. Il nécessite l'utilisation de deux images de contrastes différents parfaitement recalées. Ce modèle a été validé sur des simulations et des fantômes physique et numérique dans un premier temps. Parallèlement, la récente séquence MP2RAGE permet d'acquérir deux images co-recalées par acquisition et leur combinaison aboutit à l'obtention d'une image insensible aux inhomogénéités du signal et d'une carte de T1 des tissus imagés. Nous avons testé notre modèle sur des données in vivo MP2RAGE et avons montré que l'application du modèle linéaire de VP conduit à une sous-estimation systématique de la substance grise à l'échelle du voxel. Ces erreurs se propagent à l'estimation d'épaisseur corticale, biomarqueur très sensible aux effets de VP. Nos résultats plaident en faveur de l'hypothèse suivante : la modélisation de VP pour les images MP2RAGE doit être différente de celle employée pour des images obtenues avec des séquences plus classiques. Le modèle bi-exponentiel est une solution adaptée à cette séquence particulière. / The work developed in this thesis is within the scope of magnetic resonance imaging (MRI) acquisition and image processing for the automated analysis of brain structures. The measurement of structural modifications with time such as cortical atrophy requires the application of image processing algorithms. They must compensate for MRI artifacts such as intensity inhomogeneities or partial volume (PV) effects to allow for brain tissues segmentation then cortical thickness estimation. We suggest a new PV model relying on the physics of acquisition named bi-exponential model that differs from the commonly used linear model by modelling brain tissues and image acquisition. It requires the use of two differently contrasted and perfectly coregistered images. This model has been validated with simulations and physical and digital phantoms in a first place. In parallel, the recent MP2RAGE sequence provides two coregistered images and their combination results in a bias-field corrected image as well as a T1 map of the scanned tissues. We tested our model with in vivo MP2RAGE data and demonstrated that using the linear PV model leads to a systematic gray matter proportion underestimation in PV voxels. These errors result in cortical thickness underestimation. Our results favor the following assumption: PV modelling with MP2RAGE images must differ from the usual linear PV model applied for images obtained from more classic sequences. The bi-exponential model is an adapted solution to this particular sequence.

The inhomogeneous B1 field at higher main fields (B0) becomes more serious, leading to unsatisfactory MR image quality. To improve the signal homogeneity of routinely used T1-weighted image, usually acquired by a well-known sequence, Magnetization Prepared Rapid Acquisition Gradient Echo (MPRAGE), a new pulse sequence, Magnetization Prepared 2 Rapid Acquisition Gradient Echoes (MP2RAGE), was proposed in 2009. This technique acquires two sets of high-resolution three- dimentional images at different inversion times after a series of inversion pulses. After any of two simple calculations of the raw images (Ratio or MP2RAGE reconstruction), the output volume was obtained with dramatically reduced spatial inhomogenuity of MR signal. In this study, the contrast-to-noise ratio (CNR) optimation at 3 T was implemented independently to reproduce the previous results of other group. After that, the simulation of scanning parameters was done to optimize CNR of brain tissue at 1.5 T according to different encoding methods, different pulse sequences, and different reconstruction algorithms. Phantom and human experiments were carried on a 1.5 T scanner for further validation. The results of phantom experiment showed that both MP2RAGE and Ratio reconstructions can achiever better B1 homogeneity than MPRAGE, even with the vendor-equipped correction packages, SCIC and PURE. In addition, the agreement was made between simulation and in-vivo imaging that MP2RAGE provides higher CNR than Ratio when centric encoding also outduels linear encoding.

In recent years there has been a considerable research effort in improving the specificity of magnetic resonance imaging (MRI) techniques by employing quantitative methods. These methods offer greater reproducibility over traditional acquisitions, and hold the potential for obtaining improved information at the microstructural level. However, they typically require a longer duration for the experiments as the quantitative information is often obtained from multiple acquisitions. Here, a multi-echo extension of the MP2RAGE pulse sequence for the simultaneous mapping of T1, T2* (and magnetic susceptibility) is introduced, optimized and validated. This acquisition technique can be faster than the separate acquisition of T1 and T2*, and has the advantage of producing intrinsically co-localized maps. This is helpful in reducing the preprocessing complexity, but most importantly it removes the need for image alignment (registration) which is shown to introduce significant bias in quantitative MRI maps. One of the reasons why the knowledge of T1 and T2* is of relevance in neuroscience is because their reciprocal, R1 and R2*, have been shown to predict quantitatively myelin and iron content in ex vivo experiments using a linear model of relaxation. However, the post-mortem results cannot be applied directly to the in vivo case. Therefore, an adaptation of the linear relaxation model to the in vivo case is proposed. This is capable of predicting (with some limitations) the myelin and iron contents of the brain under in vivo conditions, by using prior knowledge from the literature to calibrate the linear coefficients. The dependence of the relaxation parameters from the biochemical composition in brain tissues is further explored with ex vivo experiments. In particular, the hyaluronan component of the extracellular matrix is investigated. The contribution to T1 and T2* is measured with a sophisticated experiments that allow for a greater control over experimental conditions compared to a typical MRI experiment. The result indicate a small but appreciable contribution of hyaluronan to the relaxation parameters. In conclusion, this work develops a method for measuring T1 and T2* maps simultaneously. These are then used to quantify myelin and iron under in vivo conditions using a linear model of relaxation. In parallel, the hyaluronan-based extracellular matrix was shown to be a marginal but measurable component in T1 and T2* relaxation maps in ex vivo experiments.