Abstract: An entirely automatic procedure for the classification of cerebral tissues from Magnetic Resonance Nuclear imaging (MRN) 3D of the head are described in this study. This procedure doesn`t make any assumption nor on the number of classes nor on the shape of the density. Indeed, this last is estimated by a non parametric method, it is about the method of the Parzen`s Kernel. A new objective function is proposed to improve the FCM algorithm by the addition of one term of entropy aiming to maximize the number of good ordering. A supplementary correction is operated by a probabilistic procedure said of fuzzy relaxation including the probabilities of the neighboring points. The validation of the algorithm is made on simulated data and on real cerebral imaging RMN.