We consider the problem of fitting the intensity distribution of an MR image using probability mass functions (PMFs) of pure and partial tissue classes. To do this, we adapt a recently proposed PMF estimation method for critically or over sampled signals. This enables us to estimate pure tissue PMFs directly from the image. Unlike most previous methods, we make no assumptions about the parametric form of the pure tissue PMFs (e.g. that they follow Gaussian distributions). We demonstrate the applicability of the algorithm to formulate partial volume segmentation in a Bayesian framework, where the a priori PMF of pure and partial tissue classes is estimated using an inequality constrained least squares method. The utility of the algorithm is shown on simulated and real data, comprised of both brain and non-brain MR images.