The ability of nonparametric models to automatically adapt to the complexity of data makes them particularly suitable for neuroimaging applications, where it is often preferable to avoid assumptions on the correct model structure. We have applied a multivariate Dirichlet process Gaussian mixture model (DPGMM) for segmenting main cerebral tissues (grey matter, white matter and cerebrospinal fluid) by learning from multiple MRI modalities (T1, T2 and PD). We experimentally show that a multivariate DPGMM produced significantly more consistent and accurate segmentations than an equivalent univariate DPGMM trained on a single modality (T1). This is also the first known attempt at performing lesion segmentation with DP mixture models. Our preliminary results show great promise, as the DPGMMs were able to correctly identify most traumatic brain injury lesions in a multimodal MRI dataset (MPRAGE, FLAIR, GE, T2 and PD) and have demonstrated their capability to learn intricate multidimensional probability distributions.