Site-Aware Longitudinal Deformation Markers from Structural MRI for Neurodegenerative Disease

Quantifying brain atrophy from serial structural MRI is central to computer-aided diagnosis but hindered by multi-site heterogeneity and the rarity of clinical conversions. We present a conversion-agnostic, deformation-based biomarker that condenses subject-level atrophy between adjacent visits. Nonlinear registration (SyN) of serial T1-weighted scans yields log-Jacobian maps aggregated into regional change features; a CN-anchored site standardization provides per-site z-scoring. We report both a raw composite and its site-normalized counterpart. We enforce \(\Delta t \ge 0.5\)  years and apply quality control. In cross-cohort neurodegeneration (ADNI: CN/MCI/AD; PPMI: CN/pPD/PD), the biomarker discriminates controls from disease groups via Mann–Whitney AUC with bootstrap CIs and Youden thresholds, tracks monotonic worsening along clinical axes, and shows limited sensitivity to site variation after normalization. The pipeline is reproducible and dataset-agnostic, offering a compact, site-normalized longitudinal readout for clinical stratification and as a foundation for generative progression modeling. Our code is available at https://github.com/erickpkgg/Longitudinal-Deformation-Biomarker.git.