Brain tissue segmentation is important in many diffusion MRI (dMRI) visualization and quantification tasks. We propose a deep learning tissue segmentation method that relies only on dMRI data. We leverage diffusion kurtosis imaging (DKI) and a recently proposed mean-kurtosis-curve (MK-curve) method to create a feature set that is highly discriminative between different types of tissues. We train a Unet model with a recently developed augmented target loss function on dMRI data from the Human Connectome Project. We show improved segmentation performance compared to several other methods and reliable segmentation results when applied on data with a different acquisition.