Deciphering Longitudinal Anatomical Representations In Brain MRI
Abstract
Purpose
Longitudinal brain MRI enables analysis of anatomical changes associated with aging and neurodegenerative disease, yet interpreting structural evolution over time remains challenging due to subtle changes such as inter-subject heterogeneity, and variability across diagnostic and cognitive states. Many existing methods rely on population-level trends or static comparisons, limiting subject-specific longitudinal analysis. The purpose of this work is to develop a framework for analyzing longitudinal brain MRI to support investigation of disease-related structural dynamics.
Methods
We propose a longitudinal modeling framework that encodes three-dimensional T1-weighted brain MRI into a multi-scale discrete latent representation using a vector-quantized variational autoencoder (VQ-VAE). Temporal relationships across scans are modeled using a latent dynamics module that enforces longitudinal ordering of representations. Subject-level variables are incorporated to enable stratified analysis of anatomical trajectories, including age, sex, diagnostic category (cognitively normal, mild cognitive impairment, Alzheimer’s disease), cognitive performance, neuropsychiatric measures, depression score, body mass index, and genetic risk factors. The framework was evaluated on longitudinal brain MRI from the ADNI dataset. To assess whether subject-level information was preserved, auxiliary inference models were trained using observed MRI data to estimate demographic, cognitive, functional, and biological variables and applied to MRI-derived representations.
Results
The proposed framework produced temporally consistent longitudinal representations that preserved region-specific structural variation as well as the conditioned covariates. Analysis of hippocampal and amygdala representations showed improved alignment with known neurodegenerative patterns compared with baseline longitudinal modeling approaches. Inferred covariates demonstrated strong agreement with reference values across age, diagnostic category, and cognitive and functional measures, indicating that the representations maintained meaningful associations with subject-level characteristics.
Conclusion
This work presents a representation-based framework for analyzing longitudinal brain MRI that supports interpretation of anatomical evolution over time and investigation of disease-related structural dynamics across heterogeneous subject populations.