Résumé:
The complexity and diversity of the methods used in neuroimaging constitute areas of innovation and discovery in science with new algorithms and new methods for the collection and analysis of brain images. They continue to open up new avenues for understanding the brain; how it is structurally and functionally organized, how it develops, and its measurements relate to knowledge, behavior, and even genetics.
The analysis of brain images allows for a deep understanding of cognitive tasks and the disease process. In particular, the analysis of anatomical MRIs in the field of computational anatomy, which is interested in the modeling of aspects relating to the structure of the brain.
The processing and manipulation of functional images cover modeling the causal relationships over time of functional data, which involves testing functional and connectivity hypotheses. The analysis of MRI images then consists of collecting and analyzing structural connectivity. The modeling of the interactions and the connectivity between the regions of the brain often calls upon the mapping mechanism and graphical calculation tools. Brain mapping approaches are increasingly used on different spatial and temporal scales. In order to understand the organization of the human brain, it becomes necessary to integrate and combine the different information modalities in order to develop a multimodal model of the brain. Atlases are a topographical means of integrating this information into a meaningful form.
In this thesis, the aim is to propose an approach based on graphs, in order to represent the different structures of the brain as well as the connectivity between them in the form of a network. It also involves adapting the analysis methods using mapping techniques between graphs. The use of graphs as a model to represent spatial and temporal information contained in brain images appears to be widespread in the field of functional/structural analysis of the brain and computational anatomy. The graph representation also seems suitable for the construction of 3D graphic models for multidimensional visual analysis of the brain. However, the algorithms for processing, learning, identification, clustering, etc..., which are often based on computational intelligence techniques, prove to be very expensive because of the complexity of the matching graph which constitutes a routine of the basis for the mapping mechanism in these algorithms.
The other aim of this thesis is to propose calculation tools that optimize the different algorithms thanks to parallelization/distribution mechanisms on high performance computing virtual machines in order to reduce the complexity of the mapping operation in the graphs ( by graph matching).
