Brain segmentation in magnetic resonance imaging (MRI) is an important step in analysis and study of the brain. While manual segmentation is the one that gives the highest accuracy, it is a labor-intensive process. Consequently, numerous automated brain segmentation algorithms for MRI have been proposed in the past. Nevertheless, none have satisfactorily addressed the entire brain extraction problem across diverse datasets in a generic and robust way. To overcome these limitations of existing methods, we propose employing UNETR for skull stripping MRI. The UNETR, has demonstrated a excellent performance in medical images analysis due to the capability of modeling long-range dependencies and capturing global context. The UNETR model was trained with 347 images from the dataset CC359, with silver standard annotation. We assessed the performance of the UNETR using a five different public datasets (LBPA40, ADNI, OASIS, NFBS and CC359), four of them never seen by the network, and compared it with six other methods considered the state of the art (BET, CONSNET, HD-BET ROBEX, PARIETAL and SynthStrip). The UNETR demonstrates a higher robustness and capability of generalization, achieving a mean dice of 0.955, staying competitive between the state-of-the-art methods.