Discovery of the image voxels of the brain that represent real activity is, in general, very difficult because of a weak signal-to-noise ratio and the presence of artifacts. The first tests of the classical data mining algorithms in this field showed low performances and weak quality of recognition. In this article, a new interactive data-driven approach to functional magnetic resonance imagery mining is presented, allowing the observation of cerebral activity. Several non-supervised classification algorithms have been developed and tested on sequences of fMRI images. The results of the tests have shown that the number of classes, signal-to-noise ratio, and volumes of activated and explored zones have a strong influence on the classifier performances.