Fig. 2

Image processing pipeline for the KN-BSI and cortical thickness computation. First, T1w scans at each timepoint were affinely registered to the baseline scan that had been rigidly registered to the ADNI template. Second, an averaged image was generated from the affinely registered scans at each timepoint. Third, T1w scans at each timepoint were affinely registered to the averaged image. DL + DiReCT was applied to the affinely registered image at each time point for segmentation and cortical thickness estimation. Segmentations for the whole brain and lateral ventricle were generated from the resultant images from DL + DiReCT. Hippocampal segmentations were generated using hippodeep. Finally, KN-BSI was computed for longitudinal volume changes in the whole brain, lateral ventricle, and hippocampus using the affinely registered images and segmentations. Abbreviations: KN-BSI, k-means normalized boundary shift integral; DL + DiReCT, deep learning-based neuroanatomical segmentation + diffeomorphic registration-based cortical thickness; tpN, Nth timepoint