Section: New Results

MEEG and Diffusion MRI

Cortical surface parcellation via dMRI using mutual nearest neighbor condition

Participants : Brahim Belaoucha, Maureen Clerc, Théodore Papadopoulo.

This work aims at parcellating the cortical surface from individual anatomy. With respact to previous works, it works for the whole brain and produces connected patches. The parcellation is obtained using the Mutual Nearest Neighbor (MNN) criterion to obtain regions with similar structural connectivity. The structural connectivity is obtained by applying a probabilistic tractography on the diffusion MRI (dMRI). Several similarity measures of connectivity are compared. The results of our method are compared to some of the atlases that can be found in the literature. We show that these atlases have lower similarity of structural connectivity than the proposed algorithm implying that the regions of the atlases may have lower functional homogeneity.

This work has been published in [27].

Iterative estimation of focal sources and their interactions constrained by dMRI

Participants : Brahim Belaoucha, Mouloud Kachouane, Théodore Papadopoulo.

This work aims at further exploiting the dMRI constraints: not only sources are constrained anatomically by patches (extracted by the method of the previous paragraph) but their dynamical behaviour is constrained by a brain network extracted from an individual dMRI. The framework reconstructs spatially localized sources from Magnetoencephalogra-phy (MEG)/Electroencephalography (EEG) using spatiotemporal constraints extracted from dMRI. The spatial reconstruction is based on our previous work on patch reconstruction [71]. The source dynamics are represented by a Multivariate Autoregressive (MAR) model whose matrix elements are constrained by the anatomical connectivity obtained from dMRI. The framework assumes that the whole brain dynamic follows a constant MAR model in a time window of interest. The source activations and the MAR model parameters are estimated iteratively. The proposed framework outperforms the two-stage approaches which have traditionally been used to estimate source interactions. Such approaches first reconstruct sources and then compute the MAR model for the localized sources. They showed good results when working in high signal-to-noise ratio (SNR) settings, but fail in detecting the true interactions when working in low SNR. Our framework iteratively refines both the reconstruction and the MAR model in two steps: sources activations are first estimated for a given MAR model and then, the MAR model is estimated for a given source reconstruction. These two steps are repeated until a stopping criterion is achieved. The work is exploratory in nature and for now focuses on simulations made with real MR data. We could confirm that accurate reconstructions and MAR models can be obtained with our method in both high and low noise levels.

This work has been published in [28], [21].