Segmentation is a fundamental problem in medical image analysis. The use of prior knowledge is often considered to address the ill-posedness of the process. Such a process consists in bringing all training examples in the same reference pose, and then building statistics. During inference, pose parameters are usually estimated first, and then one seeks a compromise between data-attraction and model-fitness with the prior model. In this paper, we propose a novel higher-order Markov Random Field (MRF) model to encode pose-invariant priors and perform 3D segmentation of challenging data. The approach encodes data support in the singleton terms that are obtained using machine learning, and prior constraints in the higher-order terms. A dual-decomposition-based inference method is used to recover the optimal solution. Promising results on challenging data involving segmentation of tissue classes of the human skeletal muscle demonstrate the potentials of the method.