Learning a model from multi-source data is a challenging and topical learning problem. Thus, generalization capacity has been proposed to deal with the domain shift (i.e. various imaging vendors, modalities, and protocols) across domains. This paper tackles the out-of-distribution generalization for prostate segmentation in MRI imaging. We propose a simple approach based on the pretraining-finetuning scheme to boost the deep neural network’s generalization to unseen data in prostate MRI segmentation. This paper introduces an objective loss that seeks to minimize cross-domain distribution by adapting Kullback–Leibler (KL) divergence. To manifest the effectiveness of our approach, we perform experiments on a multi-source public dataset for prostate MRI imaging collected from six vendors. As a result, the proposed model can yield promising cross-domains generalization capacity to unseen target domain.