Using the raw diffusion MRI signal and the von Mises-Fisher distribution for classification of Alzheimer´s disease

Diffusion MRI (dMRI) offers new signals for disease classification not available using standard anatomical MRI. However, most studies transform the raw signal to a parametric model before extracting features for classification. Here, we employ a novel method that models the signal directly to extract features for classification of Alzheimer´s disease (AD) patients versus healthy controls (HC). We studied 38 AD patients and 51 HC from the Alzheimer´s Disease Neuroimaging Initiative, and evaluated the effectiveness of two sets of features for a logistic regression classifier: (1) coefficients from a mixture of von Mises-Fisher (vMF) distributions with fixed mean directions, and (2) coefficients from a spherical harmonic (SH) expansion. We compared the classification performance for these methods with that of fractional anisotropy (FA), a popular scalar metric used in dMRI. We found vMF, SH and FA features achieved mean accuracies of 86.9%, 85.6% and 76.4% respectively, suggesting benefits of “beyond-tensor” diffusion models.