Moderating registration misalignment in voxelwise comparisons of DTI data: a performance evaluation of skeleton projection

Abstract

An evaluative methodology and five accompanying performance measures were developed to quantitatively assess the performance of the skeleton projection algorithm constituting the heart of tract-based spatial statistics (TBSS). The performance measures were designed to quantify the accuracy of skeleton projection in its indented task of alleviating any residual misalignment that may remain after image registration. A ground truth fractional anisotropy (FA) image was slightly warped using a realistic warp field that served to model post-registration residual misalignment of varying magnitudes. Skeleton projection was then used to register the warped FA image to the ground truth. Performing skeleton projection was found to yield up to 50% better correspondence between the values of FA compared to smoothing, despite the fact that less than 10% of post-registration misalignment was corrected. The align-max-with-max strategy underlying TBSS was posited as a potential explanation for this high correspondence in the values of FA, at the expense of lesser alignment between anatomically concordant voxels.