How background noise shifts eigenvectors and increases eigenvalues in DTI

Abstract

The signal-to-noise ratio of in vivo diffusion tensor imaging (DTI) is usually very limited, especially if high resolution data is acquired. In a variety of settings, the signal of diffusion weighted images can drop below the background noise level yielding an underestimated diffusion constant. In this work, we report two new artefacts in DTI that are important in this regime. Both artifacts are described analytically and numerically and are demonstrated in DTI phantoms and in subjects in vivo. First, eigenvectors are systematically shifted towards distinct 'attractive' orientations of the gradient scheme. Second, certain eigenvalues can be overestimated due to the underestimation of the measured diffusion, which can result in the misordering of eigenvalues. We show that these effects are relevant for current clinical settings of DTI.