Automatic Estimation of Error in Voxel-Based Registration

Image registration driven by similarity measures that are simple functions of voxel intensities is now widely used in medical applications. Validation of registration in general remains an unsolved problem; measurement of registration error usually requires manual intervention. This paper presents a general framework for automatically estimating the scale of spatial registration error. The error is estimated from a statistical analysis of the scale-space of a residual image constructed with the same assumptions used to choose the image similarity measure. The analysis identifies the most significant scale of voxel clusters in the residual image for a coarse estimate of error. A partial volume correction is then applied to estimate finer and sub-voxel displacements. We describe the algorithm and present the results of an evaluation on rigid-body registrations where the ground-truth error is known. Automated measures may ultimately provide a useful estimate of the scale of registration error.