Multi-scale 3D CNN with imbalance-aware training for structural MRI-based six-class Alzheimer disease staging

Alzheimer's disease staging from structural magnetic resonance imaging (MRI) is challenging when clinically adjacent stages and severe class imbalance are considered simultaneously. This study evaluates six-class classification of cognitively normal, subjective memory complaint, early mild cognitive impairment, mild cognitive impairment, late mild cognitive impairment, and Alzheimer's disease using T1-weighted MRI from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The final cohort included 3592 MRI volumes from 904 participants, with subjective memory complaint representing 3.79% of the data. We developed a parameter-compact three-dimensional convolutional neural network (3D CNN) with multi-scale volumetric feature extraction and imbalance-aware training using class weighting, focal loss with label smoothing, weighted sampling, and minority-class intensity perturbation. In the primary evaluation, partitioning was performed at the MRI ImageID (scan) level: a fixed test set of 719 MRI volumes was excluded from model development, while the remaining 2873 vol were used in five stratified training and validation folds. ImageIDs were mutually exclusive across partitions; however, participants were not grouped, and longitudinal scans from the same participant could therefore occur in different partitions. The proposed model achieved 0.9018 ± 0.0452 accuracy, 0.8804 ± 0.0393 macro F1, 0.8762 ± 0.0544 Matthews correlation coefficient (MCC), and 0.9463 ± 0.0056 macro area under the receiver operating characteristic curve (AUC) under this scan-level protocol. A same-data EfficientNet-B0 baseline achieved stronger aggregate performance, while the proposed model used 43.7% fewer parameters. Paired ablation analysis showed that weighted sampling did not improve aggregate metrics but increased subjective memory complaint precision, recall, and F1, with all corresponding 95% confidence intervals excluding zero. These findings support class-specific evaluation under severe class imbalance while distinguishing scan-level discrimination from participant-independent generalization.