Unsupervised spatio-functional clustering of fMRI data based on new functional feature

Samuel Emeriau, Frédéric Blanchard, Jean-Baptiste Poline, Laurent Pierot et Éric Bittar

As fMRI data is high dimensional, applications like connectivity studies, normalization or multivariate analyses, need to reduce data dimension while minimizing the loss of functional information. In our study we use connectivity profiles as a new functional feature to aggregate voxels into clusters. This offers two major advantages in comparison with the current clustering methods. It allows the analyst to deal with the spatial correlation of noise problem, that can lead to bad mergings in the functional domain, and it is based on the whole data independently of a priori information like the General Linear Model (GLM) regressors. We validated that the resulting clusters form a partition of the data in homogeneous regions according to both spatial and functional criteria.

Mots clés

fMRI, unsupervised clustering, feature extraction, connectivity profile, noise.