"oasis dataset"

"Early, reliable detection of Alzheimer\u2019s disease (AD) from routine structural MRI is challenging because models often overfit single sites, leak patient identity across splits, and degrade under class imbalance. We present ResidualSE\u2013Transformer, a compact hybrid that combines a Residual Squeeze-and-Excitation CNN with a  Transformer encoder to capture local anatomy and global context. We enforce patient-level partitions and first attempted four-class staging on OASIS (control, very mild, mild, moderate), to a binary stage (control vs. dementia) because the moderate class had only two subjects total, and adjacent-stage confusion dominated; the binary formulation matches clinical screening and yields stable estimates. Training uses class-weighted loss and classical augmentation. Generalization is assessed by pretraining on OASIS with fine-tuning on MIRIAD\/ADNI and by leave-one-dataset-out (LODO) testing across OASIS, MIRIAD, and ADNI. On OASIS (binary), the model achieves an accuracy of 0.791 and a weighted F1 of 0.784. LODO AUCs are 0.867 (MIRIAD), 0.535 (OASIS), and 0.503 (ADNI), with calibration consistent with these gaps. Grad-CAM highlights AD-relevant regions (medial temporal\/hippocampal band, ventricular borders, posterior cingulate) most clearly on MIRIAD. The study provides a transparent, leakage-free, calibration-aware baseline and a practical recipe for pooled multi-site training with per-site threshold calibration and patient-level aggregation for moving MRI-based AD screening toward robust cross-site deployment."