Aller au contenu principal

Soutenance de Thèse

Types d'Actualités

Le vendredi 25 janvier à 9h30, dans la Salle des Thèses du Campus de Beaulieu, Université de Rennes 1, Monsieur Wentao Xiang soutient son doctorat intitulé : " Dynamical Causal Modelling to inter changes in brain connectivity", devant le jury composé de : 

  • Corinne Mailhes, Professeur à l'Université de Toulouse
  • Catherine Marque, Professeur à l'Université de Technologie de Compiègne
  • Radu Ranta, Maître de Conférences (HDR) à l'Université de Lorraine
  • Huazhong Shu, Professeur à Southeast University
  • Ahmad Karfoul, Maître de Conférences à l'Université de Rennes 1
  • Régine Le Bouquin Jeannès, Professeur à l'Université de Rennes 1

Abstract: Our work mainly focuses on inferring effective connectivity in distant neural populations involved in epileptic seizures using a model-based technique, the spectral dynamic causal modelling (DCM). A neural mass model (NMM) is used to describe the observed epileptic intracerebral signals and their power spectral densities. DCM includes mainly two steps (i) model inversion based on the maximization of the free energy concept using the variational estimation-maximization (EM) algorithm to identify the parameters of the model and (ii) model comparison where the best model structure in terms of the maximized free energy is identified among other possible structures as the one underlying the observed data. As spectral DCM reveals some sensitivity to the initialization during the variational EM process, a misestimation of the model structure may arise. To cope with this issue, we propose two variants of spectral DCM, the L-DCM and the D-DCM algorithms. While L-DCM is based on a local adjustment of the initial guess, D-DCM relies on a deterministic annealing scheme. The performance of the proposed strategies in terms of effective connectivity inference is assessed using simulated and real human epileptic SEEG (stereoelectroencephalographic) signals. Regarding simulated and real signals, two kinds of NMM are investigated, the physiology-based model (PBM) and the complete physiology-based model (cPBM). Our experiments show the efficiency of the proposed approaches compared to the standard spectral DCM using either PBM or cPBM. The reported results also confirm that cPBM offers lower computational complexity and better estimation quality of the model parameters compared to PBM. Besides, in order to cope with the complexity of spectral DCM which is essentially related to the Gauss-Newton method used in the variational EM algorithm, a simpler ascent gradient method based on an exact line search (ELS) scheme can be employed. It allows for an optimal computation of the gradient step size to be used at each iteration towards the final solution in the given search direction. The feasibility of the ELS scheme in a probabilistic framework is not straightforward and, in this work, the ELS scheme is considered in the context of Gaussian mixture models (GMM) to accelerate the standard EM algorithm. Numerical results using both simulated and real datasets show the efficiency of the proposed ELS scheme when applied to the standard EM algorithm as well as to anti-annealing-based acceleration techniques derived from either the EM algorithm or the expectation conjugate gradient one. The ELS feasibility being proved, its applicability on spectral DCM will be an extension of the present work.