Estimating clinical significance of changes in regional brain volumetry using only two consecutive images is difficult because algorithmic measurement error dominates actual biological changes in short-term (less than a year) follow-up imaging. Here, we evaluate an approach to compute reference ranges from image pairs, using empirical Bayesian regularization. With over 6400 image pairs, we evaluate the impact of regularization strength and time between images. Regularization is essential; optimal regularization amount depends on brain region. Increased time between pairs of images improves clinical discrimination in dementia. We recommend a minimum of eight months to one year to obtain discriminative atrophy estimates.