Longitudinal FreeSurfer creates a within-subject template by rigidly registering and median-filtering longitudinal timepoints (TP). Information common to all TPs is extracted from the template for unbiased TP initialization, resulting in substantial improvements over cross-sectional processing. However, this approach is not optimal in the presence of severe atrophy or other large-scale anatomical change, which causes voxels to be filtered across tissue classes. We address this problem by introducing an enhanced longitudinal stream that deforms each TP using non-linear registration to construct the template. We demonstrate considerable increases in sensitivity to cortical thinning, without affecting test-retest reliability.