FastSurfer-Based Brain Morphometry and Machine-Learning Classification Across the Alzheimer’s Disease Spectrum

Background: This study aimed to quantitatively assess structural changes in the hippocampus, amygdala, entorhinal cortex, lateral ventricles, precuneus, and posterior cingulate using deep-learning-based FastSurfer morphometry and evaluate their contribution to classification across the Alzheimer’s disease (AD) spectrum. Methods: Three-dimensional T1-weighted MRI data from the AD Neuroimaging Initiative (ADNI) were analyzed using FastSurfer. The study included 791 participants (425 female and 366 male): 365 cognitively normal (CN), 273 with mild cognitive impairment (MCI), and 153 with AD. Morphometric measures included hippocampal and amygdala volumes, lateral ventricular volume, entorhinal cortical thickness and surface area, and precuneus and posterior cingulate thickness and folding index. Volumetric measures were normalized to estimated total intracranial volume (eTIV). Principal morphometric comparisons were additionally adjusted for age, sex, and education. Six machine-learning classifiers were evaluated using participant-level training (n = 632) and held-out test (n = 159) sets, with an additional MMSE-ablation analysis. Results: Hippocampal and amygdala volumes were significantly lower and lateral ventricular volume was significantly higher in AD than in CN (all p < 0.001), with these differences persisting after eTIV normalization and remaining significant after adjustment for age, sex, and education (all adjusted p < 0.001). MCI generally showed intermediate volumetric values between CN and AD. Mini-Mental State Examination (MMSE) scores correlated positively with total hippocampal (r = 0.495) and amygdala (r = 0.462) volumes and negatively with lateral ventricular volume (r = −0.256) (all p < 0.001). For three-class CN/MCI/AD classification, Logistic Regression with ElasticNet achieved a mean macro-F1 score of 0.706 ± 0.017 in five-fold cross-validation and a held-out macro-F1 score of 0.705 with MMSE included. Excluding MMSE reduced the held-out macro-F1 score to 0.531 and ROC-AUC to 0.735, compared with 0.857 when MMSE was included. Performance was higher for binary CN-versus-AD classification, with a held-out test macro-F1 score of 0.919 and ROC-AUC of 0.990. Conclusions: FastSurfer-based morphometry demonstrated medial temporal atrophy and lateral ventricular enlargement across the CN–MCI–AD spectrum, with the principal volumetric differences remaining robust after adjustment for age, sex, and education. The intermediate morphometric profile of MCI and the associations between medial temporal volumes and cognitive performance support the relevance of these structural measures. Machine-learning performance improved substantially when MMSE was incorporated, indicating that the combined models reflect integrated morphometric, demographic, genetic, and cognitive information rather than morphometry alone. External validation in independent cohorts is required before clinical application.