Most processing pipelines for diffusion MRI (dMRI) require an intracranial mask image to exclude voxels outside the skull, and some dMRI analyses also need a segmentation between the voxels that are primarily tissue or cerebrospinal fluid (CSF). dMRI is challenging for most segmentation methods because it usually has relatively severe image artifacts and coarse resolution. However, it does provide information about the physical properties of the material(s) in each voxel, which can be directly applied to segmentation. We describe the training of a random forest classifier to segment dMRI into intracranial, brain, and CSF masks, and compare its results to three other segmentation methods commonly used in dMRI processing. The effect of correcting smooth spatial intensity variations on dMRI segmentation is also tested.