It is easy to yield the analytical expression of the probability of false alarm with respect to the threshold, but hard to obtain the analytical expression of the threshold with respect to the probability of false alarm. By making use of the perfect properties of radial basis function (RBF) neural networks, such as approaching arbitrary non-linear mapping and quick convergence, a new scheme based modified RBF neural networks is proposed in obtaining threshold estimation. Simulation results show that the proposed scheme is of higher accuracy in threshold estimation.