A Generalizable Feature Extractor for Alzheimer's-Related Brain MRI Tasks
Researchers evaluated whether a compact, 7.18-million-parameter 3D CNN pretrained on brain-age prediction could serve as a reusable foundation backbone for Alzheimer's disease neuroimaging. By freezing the backbone and applying Low-Rank Adaptation (LoRA) with roughly 1% additional trainable parameters, the model reached an AUC of 0.964 in distinguishing dementia on ADNI, retaining an AUC of 0.871 on OASIS-3 with zero retraining. Across six clinical and anatomical benchmarks, the frozen representations also successfully estimated amyloid status and regional brain volumes from structural T1w scans, demonstrating reliable cross-cohort transfer under severe data constraints.
Why it's worth reading
It demonstrates that a lightweight 7.18M-parameter brain-age CNN fine-tuned via LoRA can reliably generalize across Alzheimer's cohorts and clinical endpoints without large-scale retraining.