Approximating the Geisser-Greenhouse sphericity estimator and its applications to diffusion tensor imaging.

The diffusion tensor imaging (DTI) protocol characterizes diffusion anisotropy locally in space, thus providing rich detail about white matter tissue structure. Although useful metrics for diffusion tensors have been defined, statistical properties of the measures have been little studied. Assuming homogeneity within a region leads to being able to apply Wishart distribution theory. First, it will be shown that common DTI metrics are simple functions of known test statistics. The average diffusion coefficient (ADC) corresponds to the trace of a Wishart, and is also described as the generalized (multivariate) variance, the average variance of the principal components. Therefore ADC has a known exact distribution (a positively weighted quadratic form in Gaussians) as well as a simple and accurate approximation (Satterthwaite) in terms of a scaled chi square. Of particular interest is that fractional anisotropy (FA) values for given regions of interest are functions of the Geisser-Greenhouse (GG) sphericity estimator. The GG sphericity estimator can be approximated well by a linear transformation of a squared beta random variable. Simulated data demonstrates that the fits work well for simulated diffusion tensors. Applying traditional density estimation techniques for a beta to histograms of FA values from a region allow representing the histogram of hundreds or thousands of values in terms of just two estimates for the beta parameters. Thus using the approximate distribution eliminates the "curse of dimensionality" for FA values. A parallel result holds for ADC.