Abstract
Early diagnosis of Alzheimer's disease is critical for effective therapeutic intervention. The progressive nature of cognitive decline requires precise computational methods to detect subtle neuroanatomical changes in prodromal stages. Current multi-scale neural networks have limited cross-scale feature integration capabilities, which constrain their effectiveness in identifying early neurodegenerative markers. This paper presents an Efficient Cross-Scale Feature-Aware Self-Attention Network (X-FASNet) designed to address these limitations through systematic hierarchical representation learning. The proposed architecture implements a dual-pathway multi-scale feature extraction approach to identify discriminative neuroanatomical patterns across various spatial resolutions, while integrating a novel cross-scale feature-aware self-attention module that enhances inter-scale information exchange and captures long-range dependencies. Quantitative evaluations on the DPC-SF dataset demonstrate that X-FASNet achieves superior performance with 93.7% accuracy and 0.973 F1-score, outperforming CONVADD by 10.8 percentage points in accuracy and 0.118 in F1-score, while also surpassing EfficientB2 on key performance metrics. Comprehensive experimentation across multiple neuroimaging datasets confirms that X-FASNet provides an effective computational framework for neurodegeneration assessment, characterized by enhanced detection of subtle anatomical variations and improved pathological pattern recognition.