The determination of the optimal architecture of a multilayer perceptron (MLP) to solve a specific problem is an important and a difficult task. Several approaches based on saliency analysis, such as the Optimal Brain Surgeon method (OBS), have been developed in this field. Nevertheless, we show in this paper that OBS does not gives an optimal architecture. We also present the advantages of applying OBS on the architecture obtained by a variable selection method. New hybrid methods are proposed. A comparison of our approaches to standard techniques for architecture optimization is presented Simulation results obtained on the Monk's problem illustrate the specificities of each method described in this paper.