“…These models can be easily deduced from the conditional distribution p(t, s, θ|y) confirming that there is no need to define the complete joint model p(t, s, y, θ) and emphasizing the rational of using a CRF approach for segmentation purpose. Moreover, the model definition in (1) induces that the conditional models p(t|s, y, θ) and p(s|t, y, θ) are MRF with energy functions denoted by H(t|s, y, θ) and H(s|t, y, θ) obtained by omiting in expression (1) the terms that do not depend on t, resp. on s. The two-stage E-step (2) and (3) requires then to compute H…”