Segmentation of MRI brain images has great clinical and academic importance. The overlap of MR intensities of different tissue types and the vast amount of thin structures in brain images make segmentation of MRI brain images difficult. In this paper, we present an EM-style geo-cuts-based segmentation method to over come these challenges. We classify the brain images into three tissue types: white matter, gray matter, and CSF. We iteratively classify the voxels and calculate the intensity profile. We use region bias and automatic seed setting combined with intensity profile induced Riemannian metrics for the classification of voxels. We then use this classification to re-estimate the intensity profile. Experimentally, our method gives very good performance on both synthetic images with ground truth segmentation and real images with the segmentation of white matter and CSF improved over the widely used EMS method.