OATAO - Open Archive Toulouse Archive Ouverte Open Access Week

Bayesian joint detection-estimation in functional MRI with automatic parcellation and functional constraints

Albughdadi, Mohanad. Bayesian joint detection-estimation in functional MRI with automatic parcellation and functional constraints. PhD, Signal, Image, Acoustique et Optimisation, Institut National Polytechnique de Toulouse, 2016

(Document in English)

PDF - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader


Brain parcellation into a number of hemodynamically homogeneous regions (parcels) is a challenging issue in fMRI analyses. An automatic inference for the parcels from the fMRI data was proposed in the framework of the joint parcellation detection estimation (JPDE) model. However, this model still requires appropriate prior information about the number of parcels and their shapes provided through an initial parcellation, which is a challenging task since it generally depends on the subject. In this thesis, we present novel approaches for hemodynamic brain parcellation. These approaches are motivated by the fact that the hemodynamic response function varies across brain regions and sessions within subjects, and even among subjects and groups. The proposed approaches belong to one of two main categories, the subject-level and group-level fMRI data analysis models. For the subjectlevel fMRI data analysis, we propose three models to automatically estimate the optimum number of parcels and their shapes directly from fMRI data. The first one is formulated as a model selection procedure added to the framework of the classical JPDE model in which we compute the free energy for the candidate models, each with different number of parcels, and then select the one that maximizes this energy. To overcome the computational intensity associated with the first approach, we propose a second method which relies on a Bayesian non-parametric model where a combination of a Dirichlet process mixture model and a hidden Markov random field is used to allow for unlimited number of parcels and then estimate the optimal one. Finally to avoid the computational complexity associated with the estimation of the interaction parameter of the Markov field in the second approach, we make use of a well known clustering algorithm (the mean shift) and embed it in the framework of the JPDE model to automatically infer the number of parcels by estimating the modes of the underlying multivariate distribution. All the proposed subject-level approaches are validated using synthetic and real data. The obtained results are consistent across approaches in terms of the detection of the elicited activity. Moreover, the second and the third approaches manage to discriminate the hemodynamic response function profiles according to different criteria such as the full width at half maximum and the time to peak. Regarding the group-level fMRI analysis, we propose two new models that are able to estimate group-level parcellation and hemodynamic response function profiles. The JPDE model is extended to allow for this group-level estimation by considering data coming from all the subjects resulting in a multisubject joint parcellation detection estimation model. However, in real data experiment, it is noticed that the smoothness of the estimated HRFs is sensitive to one of the hyperparameters. Hence, we resort to the second model that performs inter and intra subject analysis providing estimation at both the single and group-levels. A thorough comparison is conducted between the two models at the group-level where the results are coherent. At the subject-level, a comparison is conducted between the proposed inter and intra subject analysis model and the JPDE one. This comparison indicates that the HRF estimates using our proposed model are more accurate as they are closer to the canonical HRF shape in the right motor cortex. Finally, the estimation of the unknown variables, the parameters and the hyperparameters in all of the proposed approaches is addressed from a Bayesian point of view using a variational expectation maximization strategy.

Item Type:PhD Thesis
Uncontrolled Keywords:
Institution:Université de Toulouse > Institut National Polytechnique de Toulouse - Toulouse INP (FRANCE)
Laboratory name:
Research Director:
Tourneret, Jean-Yves and Chaari, Lotfi
Deposited On:10 Oct 2016 09:50

Repository Staff Only: item control page