New local trabecular support ranks formation-labelled transitions: An image-derived framework for bone adaptation in human distal-tibia based on HR-pQCT.

Abstract

Longitudinal HR-pQCT scans can identify mineralised-tissue gains and losses after image registration, but whether baseline trabecular architecture can rank the locations of subsequent image-derived transitions remains unclear. This study tested whether surface-conditioned local trabecular support can rank subsequent formation-labelled transitions directly from baseline images and whether a longitudinal axial-prolate neighbourhood provides additional information beyond the isotropic local support. Baseline and 6-month distal-tibia HR-pQCT scans of nine postmenopausal women are analysed, giving 18 side-level image pairs. The analysis is restricted to baseline surface-marrow voxels, defined as marrow voxels sharing a face with baseline bone. A spherical support count provides an isotropic local-BV/TV-equivalent baseline. The proposed axial-prolate operator counts baseline bone voxels within an ellipsoidal neighbourhood, with its long axis aligned with the scanner's z-axis. Ranking performance is evaluated using leave-one-subject-out ROC-AUC. The obtained results indicate that the spherical local support ranks subsequent formation-labelled transitions above chance, showing that an isotropic local bone amount already contains prospective surface-conditioned information. The axial-prolate operator further improves this process. Matched-volume comparisons show that both anisotropy and longitudinal orientation contribute to this gain, while local-BV/TV-stratified analyses indicate that the directional signal is not fully reducible to the isotropic local bone amount. These findings establish an image-derived framework for ranking trabecular formation-labelled transitions from the baseline architecture alone. The distal-tibia application indicates that site-informed directional kernels can refine the general local-support framework. External validation is required to determine whether the kernel scale and alignment should be recalibrated for other anatomical sites and populations.