Predicting early mild cognitive impairment utilizing transformer-based multiscale attention with across-space and time mechanism

Alzheimer's disease (AD), a progressive and irreversible brain disorder, is the leading cause of dementia. Mild Cognitive Impairment (MCI), being the transitional phase between cognitively normal (CN) and AD, makes early identification of its progression critical for timely therapeutic intervention. Magnetic Resonance Image (MRI) can detect changes in brain regions and has been widely used in AD diagnostic studies. Current diagnostic paradigms predominantly rely on static single-timepoint MRI analysis through deep learning, yet critically overlook the temporal evolution of neuropathological features and the hierarchical changes of brain structural. This paper proposes a Transformer-based framework that synergistically integrate multi-scale MRI feature extraction with longitudinal pattern analysis. The architecture employs CNN for initial feature extraction, coupled with novel multi-scale attention to establish long-range dependencies and fuse related features over time. Tested on the ADNI dataset, the model achieved 80.40% accuracy and an AUC of 0.8395 for sMCI and pMCI classification, outperforming other MCI prediction studies.