Predicting the progression of mild cognitive impairment based on fine-grained and spatiotemporal features of MRI

Abstract

Alzheimer’s disease is a significant neurodegenerative disorder that severely impairs cognitive function and behavior, making it the predominant cause of dementia worldwide. As a precursor state to AD, early diagnosis of Mild Cognitive Impairment (MCI) is critical for timely intervention and treatment. Current single time-point structural MRI (sMRI) analysis methods inadequately capture temporal evolution of neuropathological features and the hierarchical changes of brain structural. To address this limitation, we proposed MME-TransENet, a novel hybrid CNN-Transformer architecture, predicting the progression of MCI based on fine-grained and spatiotemporal features of MRI. The architecture synthesizes longitudinal sMRI data from dual temporal points and integrates three novel components: EFF-GCNet for local feature extraction capturing global dependencies, multi-scale attention mechanism modeling spatiotemporal feature interactions, and adaptive fusion of fine-grained pathological signatures. Evaluated on the ADNI dataset, MME-TransENet achieves state-of-the-art performance (84.74% accuracy, 0.8587 AUC), outperforming existing methods. Visualizations via Grad-CAM confirmed the focus of MME-TransENet on clinically critical regions and demonstrated its ability to learn spatiotemporal patterns of biomarker. This study demonstrates that longitudinal feature learning through multi-scale integration significantly enhances MCI diagnostic precision, offering a robust method for tracking neurodegenerative progression.