Exploring Whole Brain fMRI Data with Unsupervised Artificial Neural Networks

fMRI data analysis is challenging because data are generally a modest signal embedded in a high dimensional space. We present here a method for classification and discrimination among fMRI data that is based on the Self Organizing Map (SOM) algorithm. This allows no prior selection of spatial or temporal features. We applied the method to single-subject and inter-subject classification. fMRI data are issued from a block design experiment where subjects were passively viewing emotional pictures of positive, neutral and negative valences. These data are classified with an unsupervised non-linear method, the SOM. We demonstrate here that the SOM algorithm is a good candidate for multi-voxel pattern analysis methods as it leads to good performance and it allows to extract information about cognitive processes. Our method presents three phases: data dimensionality reduction: where non relevant data for classification (ie non grey matter) are deleted from a volume, SOM algorithm training: where statistic regularities relevant for classification are extracted, SOM algorithm test: where classification and recognition performance are calculated.