Analyse de l`angiogénèse tumorale par DCE imagerie
Transcription
Analyse de l`angiogénèse tumorale par DCE imagerie
Analyse de l’angiogénèse tumorale par DCE imagerie : Développement de nouveaux outils pour améliorer la quantification (see english version on third page ) Thesis Supervisors: Professor Nikos Paragios ([email protected]), PhD/D.Sc. & Professor Charles-Andre Cuenod, PhD/D.Sc./MD École Centrale Paris / Hôpital Européen Georges Pompidou – LRI U970 PARCC HEGP Abstract : La prise en charge du cancer est un des problèmes majeurs actuellement. Les techniques d’imagerie à des fins diagnostique et de suivi thérapeutique se sont récemment renforcées par l’utilisation de techniques fonctionnelles qui permettent d’adjoindre aux critères purement morphologiques des critères physiologiques qui sont réputés pour être plus sensibles et pour se modifier plus précocement. Cela ouvre la perspective de diagnostics plus précoces et de suivis thérapeutiques plus précis. Parmi ces techniques fonctionnelles, les techniques d’imagerie dynamique avec injection d’agent de contraste (Dynamic Contrast Enhanced Imaging) sous IRM (DCE-MRI), scanner (DCE-CT) ou échographie (DCE-US) sont les plus fréquemment utilisées. Les résultats obtenus par ces techniques devraient pouvoir servir de « biomarqueurs » pour suivre les patients et évaluer les nouveaux traitements antitumoraux. Ces techniques néanmoins souffrent, en conditions cliniques, de limitations qui les rendent peu reproductibles et peu fiables. Les limitations de ces techniques basées sur l’analyse de séquences temporelles sont essentiellement due à trois facteurs : les mouvements physiologiques (respiration, battements cardiaques, mouvements involontaires …), le faible rapport sur bruit des images individuelles (du fait de temps de poses courts rendu nécessaire pour l’acquisition dynamique) et la simultanée temporelle des différentes cinétiques qui peuvent se superposer (perfusion, passage trans-capillaire et diffusion interstielle) et entraîner une indétermination avec solutions multiples de l’extraction des paramètres de microcirculation. Objectif Le but général de ce projet est de construire un « pipeline » complet pour la quantification automatique de l’angiogénèse tumorale grâce à l’imagerie dynamique de contraste (DCE-Imaging). L’approche proposée devra remplacer les critères morphologiques (déterminés manuellement) et très critiques qu’est l’évaluation RECIST, par des biomarqueurs d’imagerie qui exploitent la nature 3D des tumeurs et leurs dynamiques fonctionnelles. Afin de répondre à l’objectif clinique décrit ci-dessus, nous devrons introduire de nouvelles méthodes afin de résoudre les deux problèmes les plus fondamentaux en imagerie biomédicale que sont : la fusion des images biomédicales, qui consiste à déterminer une méthode de recalage dense capable de tenir compte des mouvements dus à la respiration, aux battements cardiaques et aux mouvements des patients. Les méthodes de recalage conventionnelles échouent dans ce domaine du fait même de la nature dynamique du rehaussement des tissus qui induisent des modifications dans l’apparence même des images. La segmentation des images biomédicales qui a comme but de regrouper les voxels de tissus ayant un comportement fonctionnel similaire, notamment pour améliorer la qualité de l’analyse des différents comportements et pour fournir une détection automatique des tumeurs ou d’autre tissus comme la fibrose post-thérapeutique ayant une cohérence de mesures dans le domaine spatiotemporel. Programme de Recherche Ce projet consiste en trois axes scientifiques majeurs : la composante de fusion d’images déformables, la composante de détection automatique et de segmentation des tissus tumoraux, et la composante de modélisation fonctionnelle permettant d’extraire les nouveaux biomarqueurs. - Mois 1-18 : La composante de fusion d’image déformable sera traitée par l’utilisation de modèles graphiques et d’optimisation discrète. L’idée centrale sera de représenter l’estimation des déformations en 3D comme un problème de recherche de la solution de coût minimum dans un graphe. Le challenge scientifique le plus important sera de considérer le problème de fusion simultanée de tous les volumes acquis à chaque temps et de définir une métrique adéquate. - Mois 6-24 : Le problème de détection spatiotemporelle des voxels tumoraux et de leur regroupement sera traité par l’agglomération par clusterisation automatique des signaux temporaux de rehaussement. L’idée centrale est de procéder à une hiérarchisation des comportements et des regroupements statistiques. - Mois 12-30 : La modélisation des données dynamiques obtenues in vivo, ne pourra se concevoir qu’après les deux étapes précédentes afin de travailler sur des données d’origine non univoque et corrigés du bruit induit par les mouvements. L’idée principale est de déconvoluer les rehaussements tissulaire par la fonction d’entrée artérielle en adaptant automatiquement de façon récursive le model le plus adapté aux donnés à l’aide de critère de qualité. Cette récursivité incluant à terme le mode d’injection de l’agent de contraste en fonction du type tissulaire voire à l’utilisation d’agents de contrastes dédiés de poids moléculaires variables. Le but final de la modélisation est l’extraction des biomarqueurs microcirculatoires. - Mois 12-30 : Evaluation clinique des biomarqueurs microcirculatoires sur la base de données de l’étude REMISCAN (étude STIC financée par l’INCA) comportant des acquisitions en DCE-CT, DCEMRI et IRM de diffusion durant le suivi de patients mis sous traitement antiangiogénique pour métastases de cancer du rein. Les données sont obtenus en multicentriques dans 16 centres hospitaliers français de façon prospective, incluant les acquisitions avant traitement, à J7 après le début du traitement et lors du suivi jusqu’à progression tumorale. Ces données sont couplées à l’acquisition morphologique classique (scanner thoraco-abdomino-pelvien) permettant l’évaluation selon les critères RECISTS. La robustesse et pertinence des biomarqueurs de microcirculation seront testées initialement sur cette base. Ultérieurement, les donnés du PHRC Myélome piloté par l’équipe de l’hôpital Henri Mondor pourront également être testées. Résultats attendus Solutions techniques pour répondre aux deux problèmes les plus cruciaux de l’imagerie biomédicale que sont la fusion d’images déformables et la segmentation. Les solutions à ces problèmes pourront avoir de très nombreuses applications dans le futur et un important impact clinique. Développement de biomarqueurs d’imagerie de référence pour l’évaluation et le suivis des tumeurs, biomarqueurs acquis de façon non invasive et extraits de façon entièrement automatisées grâce à l’imagerie dynamique avec injection. Procedural Modeling of architectures towards large-scale image-based modeling of urban environments Thesis Supervisor: Professor Nikos Paragios ([email protected]), PhD/D.Sc., École Centrale Paris Abstract: The aim of this thesis is to provide a novel, efficient, compact and scalable method for image-based automatic representation and large-scale modeling of urban scenes. To this end, we plan to combine efficient representations derived from shape grammars and procedural modeling with powerful image-based optimization techniques towards automatic 3D parsing of urban environments from a limited number of 2D images (google street-view like for example). Such an approach offers compactness (buildings are represented using strings corresponding to the grammar derivation rules), modularity (different architectures can be expressed through simple modifications of the grammar rules and its final shapes), scalability (grammars are very powerful representations with limited complexity and therefore can easily scale) & computational efficiency (grammar-like structured optimization can be done fast and with certain optimality guarantees leading to very promising results) while at the same times offers a complete contextual representation of the urban environment (3D models refer to urban representations where all architectural elements are detected and contextually represented). Problem Definition / Motivation / Prior Art 3D large-scale urban modeling has became a major research direction in the recent years due to the development of the computer games industry, urban planning, navigation & other commercial services, etc. Existing method exploit a variety of sensors like for example digital elevation maps, range scanners, aerial imaging and binocular vision cameras towards producing 3D urban models. Image-based methods which inherit reduced acquisition cost employ conventional geometric techniques towards recovering surface representation of the scene. The central idea is that if a 3D point is visible to more than one image and one can determine the exact position of this point in two different images, then the relative depth of this point can be determined. Recent results on geometric reconstruction [1] have shown that these methods can now be used for modeling large scale environments towards producing “triangle-based” representations. The main limitations of these methods refer to their inability to provide a structured representation of the scene, the sparseness of the obtained 3D models and their lack of compactness/interpretability (point clouds maps). Shape grammars [2] is a geometric modeling and interpretation schema that aims at expressing complex scenes through a syntax like representation that consists of a sequence of rules and a set of final shapes corresponding to rules that cannot be derived further. This concept is inspired from conventional grammars, with phonemes being replaced with geometric shapes and syntax rules from shape composition rulers (as shown in Fig. (1)). Procedural modeling (inspired from shape grammars) has gained significant attention in the field of computer graphics the past decade [3] towards automatic generation of urban environments with primary application field the domain of computer games. The main challenge of such an approach relates to the definition of the grammar which in most of the cases was done mostly using the end-user expertise (derivation rules, final shapes, etc.). The use of these concepts to building modeling and reconstruction has gained significant attention recently, (like for example in [4]). In this context reconstruction can be reformulated as a grammar derivation problem such that the generated model expresses optimally the observed image according to a similarity criterion. Fig. (1) : Grammar derivation sequence for buildings corresponding the Haussmann's architecture. The transition between 3D modeling and 2D images, the definition of the metric and the optimization of the grammar derivation sequence can be very challenging when aiming large scale reconstruction where the camera view point and the geometric/photometric variability of buildings vary significantly. The grammar optimization is challenging due to the mixed nature of unknown variables since neither the derivation sequence is known (discrete sequence), nor the corresponding parameters of each rule (continuous variables). The case of fixed grammar was addressed in [8] through discrete optimization and in [10] through gradient descent. Reversible-jump Markov chain Monte Carlo allows simulation of the posterior distribution on spaces of varying dimensions and has been used to optimize grammars for building reconstruction in [5]. The same principle was applied in [6] using multiple hypothesis testing while constraining the maximum derivation tree towards computational efficiency while graphbased optimization methods were considered in [7]. In terms of image-driven cost, the use of geometric proximity is often considered when range data are available [9] while image structure (selfsimilarities) was introduced in [8], mutual information in [10] and automatic [6] or user-based image classification [7]. Despite the great promise of the above mentioned methods, they inherit two important limitations: • The definition of the grammar both in terms of rules, as well as in terms of final shapes is user specific, an approach that lacks of modularity, and scalability. One can overcome this limitation using generic grammars which inevitably make their inference from images almost impossible. • The image-based grammar inference assumes the camera acquisition parameters either to be known or considers only 2D grammars, while being computationally and performance-wise inefficient (convergence to local minima) making the whole processing chain almost not useful in practice. Research Goals In this thesis , following our work in grammar-based large scale urban modeling [6, 7, 8] we would like to address the above mentioned limitations. [Work Package 1] In particular, we would like to study the problem of automatic grammar inference for typologies of architectures from generic purpose grammars. Let us consider without loss of generality that a number of buildings corresponding to the same class of architecture have been modeled and reconstructed using a generic grammar. This can be done either with the help of the user or automatically using one of the methods being developed in [4-10]. The problem of deriving a representative grammar for an architecture typology can be viewed as a problem of reasoning over the structure of various graphs representing the different grammar derivation sequences of the buildings of the training class. In such a context, one can omit the derivation parameters and focus only to the derivation sequence. Ideally one would expect that in such a sequence, sub-trees will be repeated across buildings. Such sequences will determine super-rules in the new grammar and will help to customized generic grammars to architecture specific ones. In order to address such a concept, we should be able to perform graph-matching with partial correspondences, while taking into account the nature of interaction between derivation sequences that is usually higher order. To this end, our research will be based on discrete-optimization methods that do explore dual-decomposition [11]. The idea is that each sub-tree can be considered as a higher-order clique and one seeks optimal matching for the entire tree and the individual sub-trees. Due to the compact representation offered by the grammar, the inclusion of the process of almost all possible trees would be feasible providing guarantees on the optimality of the process. Once such a matching has been established, the conventional dimensionality reduction techniques based on graph embeddings can be considered to determine architecture-specific grammars. [Work Package 2] The second aspect of our work will be dedicated to pose-invariant image-based grammar inference from un-calibrated images. In this context, one seeks the optimal 3D configuration of the grammar and the corresponding image projections such that the distance between the reconstructed model and its image variant is minimized. Reinforcement Learning comprises a set of techniques to approximately, but efficiently solve problems that are phrased as Markov Decision Processes, including Dynamic Programming (DP) and Monte-Carlo as extreme cases. In order to address the burden of computational complexity, state aggregation methods will be studied that consist of a finite discretization of the state space into a finite collection of cells where cells do aggregate the states that fall in this cell. Grammar parsing than can be solved efficiently in a reduced state space regular techniques. The exploration of cell is critical in such context and if properly designed could lead to asymptotically converge to the optimal policy, at least under idealized conditions. The outcome of the automatic grammar inference for typologies of architectures from generic purpose grammars could provide a natural selection methods regarding optimal exploration of rules such that cells are visited according to their associated image confidence [being learned at the training stage]. We are convinced that the proposed approach carries great scientific innovation and could have a tremendous application impact in a number of domains, such that image-tagging, navigation, urban modeling, etc. References [1] Agarwal, S. and Furukawa, Y. and Snavely, N. and Curless, B. and Seitz, S. and Szeliski, R.: Reconstructing Rome. IEEE Computer 43(6): 40-47 (2010) [2] Stiny, G. and Gips, J.: Shape grammars and the generative specification of painting and sculpture. North Holland Publishing Co (1971) [3] Müller, P. and Wonka, P. and Haegler, S. and Ulmer, A. and Van Gool, L.: Procedural Modeling of Buildings, Proceedings of ACM SIGGRAPH / ACM Transactions on Graphics: 25(3) 614—623 (2006) [4] Vanegas, C. And Aliaga, D. and Benes, B.: Building Reconstruction using Manhattan-World Grammars. IEEE Conference in Computer Vision and Pattern Recognition: pp.358-365, (2010) [5] Alegre, F. and Dellaert, F: A probabilistic approach to the semantic interpretation of building facades. In Workshop on Vision Techniques Applied to the Rehabilitation of City Centres (2004). [6] Simon, L. and Teboul, O. and Koutsourakis, P. and Paragios, N.: Random Exploration of the Procedural Space for Single-View 3D Modeling of Buildings. International Journal of Computer Vision (2011) [7] Teboul, O. and Simon, L. and Koutsourakis, P. and Paragios, N.: Segmentation of building facades using procedural shape priors. IEEE Conference in Computer Vision and Pattern Recognition: pp. 3105-3112, (2010) [8] Koutsourakis, P. and Simon, L. and Teboul, O. and Tziritas, G. and Nikos Paragios, N.: Single View Reconstruction Using Shape Grammars for Urban Environments. IEEE International Conference in Computer Vision: pp. 1795 - 1802, (2009) [9] Toshev, A. and Mordohai, P. and Taskar, B.: Detection and Parsing Architecture at City Scale from Range Data. IEEE Conference in Computer Vision and Pattern Recognition: pp. 398-405, (2010) [10] Muller, P. and Zeng, G. and Wonka, P. and Van Gool, L.: Image-based procedural modeling of facades. ACM Transactions on Graphics: pp. 18-26, (2007). [11] Komodakis, N. and Paragios, N. and Tziritas, G.: MRF Energy Minimization and Beyond via Dual Decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(3): 531-552 (2011) [12] Sutton, R. and Barto, A.: Reinforcement Learning: An Introduction. MIT Press (1998).