UNSUPERVISED CLASSIFICATION OF WHOLE BRAIN fMRI DATA WITH ARTIFICIAL NEURAL NETWORKS

In the present study, we apply the Self Organizing Map (SOM) algorithm for classifying cognitive states from fMRI data without prior selection of spatial or temporal features. In addition, we compare our method with two other models. We applied the method to single-subject as well as multi-subject classification. BOLD signals from subjects viewing emotional pictures of positive, neutral and negative valences were acquired during a block design experiment, and classified with an unsupervised non-linear method, the SOM. We demonstrate here that, in terms of classification performance, the SOM algorithm outperforms an SVM algorithm when processing whole brain data, and performs as well as methods (SVM and KCCA) working with temporal compression or spatial feature selection. Our method presents three phases: data dimensionality reduction : where non functional data are deleted, SOM algorithm training : where statistic regularities relevant for classification are extracted, SOM algorithm test : where the subject’s brain state is predicted from his brain activity.