Unsupervised Domain Adaptation (UDA) methods have emerged as a promising solution to generalize a learning to close datasets (domains) without the need to produce new ground truth. Nonetheless in biomedical images, some high domain shifts between source and target images lead to poor adaptation. To address this issue, we propose a method relying on two main ideas. First, we learn the source posterior label distribution with diffusion models. Assuming the target label distribution is similar, this learning helps us to guide the diffusion process to generate relevant segmentation masks on target domain. Alongside this probabilistic constraint, we propose a reconstruction pretext task on both source and target domain to extract common images features. Our approach is compared to the state of the art on three highly shifted mitochondria segmentation datasets. Our method ranks among the best in moderately difficult adaptation cases and succeeds in difficult adaptation cases where all other tested methods fail. Code will be available.