Estimation of the prior distribution of ground truth in the STAPLE algorithm: an empirical bayesian approach