Dealing with uncertainty in diffusion tensor MR data

This paper explains how radio frequency (RF) background noise produces uncertainty in measured diffusion tensor MRI (DT-MRI) data, how this noise can be modeled, and how its effects can be mitigated. DT-MRI data are derived from a series of magnitude diffusion-weighted images (DWI) in which RF noise is rectified. A new Gaussian distribution is proposed that describes the variability of the estimated diffusion tensor, D, in an ideal experiment in which RF noise is the only artifact present. We show how to improve the design of DT-MRI experiments by requiring that the statistical distribution of D be independent of the laboratory coordinate system. Non-parametric empirical methods of analyzing uncertainty in DT-MRI experiments are also described. Monte Carlo simulations are useful in designing and interpreting DT-MRI experiments. Bootstrap methods help us measure the true variability of D (and quantities derived from it), and assess the quality of DT-MRI data. Matrix Perturbation techniques predict how the uncertainty in D propagates to its eigenvalues and eigenvectors. A method for obtaining a continuous diffusion tensor field from the measured discrete noisy DT-MRI data also reduces the uncertainty of D and quantities derived from it. Finally, we describe schemes that use wavelets to remove noise from DWI and DT-MRI data while preserving boundaries between different tissue regions. Collectively, these parametric and nonparametric methods provide a unified statistical framework to improve the design of DT-MRI experiments and their subsequent analysis.