This paper presents an unsupervised approach for medical volume restoration. To cope with various scanning modalities and strongly corrupted data, an original information tool is introduced: the entropic deviation. To validate the robustness of this estimation, a non-linear restoration filter based on Markov random fields is proposed. No parameter tuning is required from the user thanks to the adaptive value of the entropy power. Finally, the good quality of the filtered volumes are promising for any clustering application aiming at anatomical structures extraction in medical volume datasets.