In diffusion MRI, standard approaches for fibertract identification are based on algorithms that generate lines of coherent diffusion, currently known as tractography. A tract is then identified as a set of such lines selected on some criteria. In the present study, we investigate whether fibertract identification can be formulated as a segmentation task that recognizes a fibertract as a region where diffusion is intense and coherent. Indeed, we show that it is possible to segment efficiently well-known fibertracts with classical image processing methods provided that the problem is formulated in a five-dimensional space of position and orientation. As an example, we choose to adapt to this newly defined high-dimensional non-Euclidean space, called position orientation space, an algorithm based on the hidden Markov random field framework. Structures such as the cerebellar peduncles, corticospinal tract, association bundles can be identified and represented in three dimensions by a back projection technique similar to maximum intensity projection. Potential advantages and drawbacks as compared to classical tractography are discussed; for example, it appears that our formulation handles naturally crossing tracts and is not biased by human intervention.