Anderson et al.'s GNN maps local brain aging at 1.37mm resolution, outperforming prior methods

A new graph neural network (GNN) from researchers at USC (Anderson et al., 2026) brings unprecedented spatial resolution to brain age estimation, moving from global averages to local cortical mapping. The model uses morphometric features — cortical thickness, surface area, curvature, gray/white matter intensity ratio (GWR), and sulcal depth — extracted from T1-weighted MRIs, processed on cortical surface meshes with an average inter-vertex distance of just 1.37mm. Trained on a large cohort of 14,423 cognitively normal adults (from multiple public datasets), the GNN achieves lower mean absolute error (MAE) than existing state-of-the-art methods while producing biologically plausible aging patterns.

When applied to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, the model reveals distinct spatial signatures: association cortices age fastest in cognitively normal individuals, while mild cognitive impairment shows widespread aging concentrated in the parahippocampal gyrus. Alzheimer's disease subjects exhibit significant aging across the entire cortex, particularly in medial temporal regions and connected networks. Feature ablation highlights that cortical curvature and GWR are most sensitive to AD pathology. Regional local brain age gaps correlate strongly with neuropsychological measures of cognitive impairment, directly linking imaging-based aging maps to clinical outcomes. The work provides an interpretable, high-resolution tool for studying brain aging and Alzheimer's disease progression, with code and supplementary tables publicly available.