Abstract (endast på engelska)

Diffusion weighted imaging (DWI) (Le Bihan et al. 2001) has since its inception been a tremendous tool for studying white matter in the human brain in vivo. The acquired DWI data are often modeled by a single-diffusion tensor model assuming that the magnitude of the magnetic resonance imaging (MRI) signal follows a Gaussian distribution. However, under the reasonable assumption of complex Gaussian distributed noise the magnitude of the signal follows a Rician distribution (Gudbjartsson and Patz, 1995).

The Gaussian model is a good approximation to the Rician model in functional MRI where the signal-to-noise (SNR) is almost always sufficiently high, but in diffusion tensor imaging (DTI) low SNRs are common (Zhu et al. 2009). Using the Gaussian model for DTI can therefore lead to severely misleading inferences. We introduce a new Bayesian Rician regression model to compute the posterior distribution of a single-diffusion tensor in each voxel (1.5x1.5x1.5mm^3) of the white brain matter.

Using data from the Human Connectome Project, we compare our Rician regression model to the more conventional Gaussian model. The results suggest that the Rician model gives higher Fractional Anisotropy (FA) values and, in general, a clearer pattern of diffusion compared to the Gaussian model.