Diffusion Tensor Imaging (DTI) is a modality of Magnetic Resonance Imaging (MRI) used to measure water diffusion in tissue, modeled in the form of a symmetric second rank tensor. Two scalar values derived from the tensor, Apparent Diffusion Coefficient (ADC) and Fractional Anisotropy (FA) are indicative of the mean diffusivity and the degree of anisotropy respectively. DTI has proven to be particularly useful to non-invasively track white matter (WM) fibers in the brain. The microarchitecture of the WM fibers preferentially restrict diffusion in a plane perpendicular to the fiber long axis; diffusion is therefore anisotropic. DTI has thus emerged as a powerful tool for detecting abnormalities known to be associated with loss of white matter integrity. The ultimate power of DTI is in identifying white matter tract related changes (diffusion indices along the fiber tracts, morphology of fiber tracts). However, tract based analysis is currently limited by fiber tracking techniques, variability in the placement of regions of interest for tract delineation, difficulty of fiber tracking in disease states, and availability of a limited number of fiber atlases. Probabilistic fiber atlases based on scalar indices and affine registration that have been reported may be limited in their potential to identify differences at the brain periphery where there is greater variability in the fiber tracts. The development and evaluation of probabilistic fiber atlases leveraging new tensor based non-linear registration will enable more investigations based on fiber tract based differences in normal and disease cohorts. The purpose of this study is to quantify and compare the accuracy of three different methods for probabilistic fiber atlas creation. The first method uses full tensor information to create a linearly transformed tensor atlas by affine aligning the subjects of a population to an averaged template. The second method uses the FA and ADC maps derived from the affine registered atlas. This algorithm non-linearly registers the subjects to the averaged template using information from both the subjects' FA and ADC maps to create a highly non-linear registered scalar atlas. The third method creates a non-linearly registered tensor atlas by nonlinearly aligning all the subjects to the averaged template using a piece-wise affine algorithm. Regions of interest are transferred from a labeled atlas (JHU-DTI) by appropriate affine/ nonlinear transforms. Fibers are generated in the transformed space for the affine and non-linear tensor techniques. For the non-linear scalar approach, fibers are generated in the affine transformed data and then mapped to the reference space using the deformation fields obtained in the atlas step. Probabilistic callosal white matter fiber atlases are derived from the individual subject binarized fiber volumes. Probabilistic maps are derived based on the presence/absence of fiber at a given voxel without using the fiber density information. Additional attempts to capture fiber density to form a weighted probabilistic white matter fiber maps are also proposed. Atlases created from all three methods are evaluated will be evaluated (i) visually by comparing the atlases at the same anatomic locations using color maps and surface plots, (ii) quantitatively by whole brain volume histogram distributions of probabilities, and (iii) slice by slice histograms to evaluate the accuracy as a function of the anatomical location of the tracts. Potential future applications of these atlases will be presented.