Neuroinfinity: A Lightweight Prompt-Driven Vision–Language Framework for Alzheimer’s Disease Stage Classification Using Magnetic Resonance Imaging

Alzheimer’s disease (AD) staging via routine MRI remains challenging due to subtle morphology, class imbalance, and variability in clinical phrasing. This work presents NeuroInfinity, a lightweight prompt-driven vision–language framework that reformulates AD staging as image–text alignment on 2D T1-weighted MRI slices via a frozen CLIP ViT-B/32 encoder with minimal trainable prompt adapters. In the primary cohort, NeuroInfinity attained 96.1% accuracy, 94.1% macro-F1, and 0.983 AUC, with strong calibration (balanced accuracy of 94.2%, MCC of 0.91, ECE of 3.4%, Brier of 0.049). A targeted prompt-diversification strategy notably improves the underrepresented moderate impairment class (MoI F1 91.0%), and the model exhibits zero-shot robustness to unseen clinical phrasings (94.8% accuracy, 92.7% F1). External validation on four independent public cohorts (ImagesOASIS, ADNI, and MIRIAD) demonstrated consistent generalizability (accuracy of 94.4–95.3%) with stable calibration profiles. Designed for practical use, the framework runs at ~14 ms per slice on a mid-tier GPU and maintains linear throughput scaling. These results indicate that prompt-guided image–text alignment can deliver accurate, calibrated, and interpretable AD staging under realistic data constraints while remaining computationally efficient and extensible to prospective, multisite evaluation.