Iterative reweighted linear least squares for the robust estimation of diffusion magnetic resonance parameters


Diffusion weighted magnetic resonance (DW-MR) imaging suffers from physiological noise such as artifacts caused by motion or system instabilities. This obviates the need for robust diffusion parameter estimation techniques. In the past, several techniques have been presented including RESTORE and iRESTORE. However, these techniques are based on nonlinear estimators, and are consequently computationally intensive. We present a new robust, iteratively reweighted linear least squares (IRLLS) estimator. IRLLS performs a voxel-wise identification of outliers in DW-MR images where it exploits the natural skewness of the data distribution to become more sensitive to both signal hyperintensities and signal dropouts. When compared to RESTORE and iRESTORE, IRLLS shows no significant loss in accuracy or precision, yet proves to be significantly faster. Moreover, IRLLS appears to be even more robust when considering overestimation of the noise level and with low SNR. The significant shortened calculation time in combination with the increased robustness make IRLLS a practical and reliable alternative to current state-ot-the-art outlier detection techniques for the robust estimation of DW-MR parameters.