In this paper a novel procedure for training radial basis function (RBF) networks in the presence of censored data is presented. The proposed technique is based on a decomposition determined by decoupling parameters which are estimated by maximum likelihood. The censorship considers that some outputs are missing, but classification intervals containing them are observed. Convergence of the algorithm is proved by showing that it can be framed as a GEM (generalized expectation-maximization)-based training method. Hence, the possibility to adapt a GEM algorithm to deal with censored data without assuming known error variances is proved. The robustness of the algorithm is illustrated via a simulation example.