3D-HAFNet: A Novel Hierarchical Aggregation Feature Network With Triple Attention for Early Alzheimer’s Disease Detection From Structural MRI

Abstract

Early detection of Alzheimer's disease (AD) during the mild cognitive impairment (MCI) stage remains clinically challenging despite its critical importance for therapeutic intervention. We present 3D-HAFNet, a deep learning framework for early AD detection using structural magnetic resonance imaging (sMRI) that addresses two fundamental challenges: capturing multi-scale pathological patterns and exploiting directional information inherent in anisotropic brain data. The model incorporates two key innovations: (1) a Hierarchical Aggregation Feature Pyramid (HAFP) enabling bidirectional multi-scale feature fusion across network layers, and (2) a Triple Attention (TA) mechanism extending traditional channel and spatial attention with directional attention tailored for 3D neuroimaging. Experiments on 1247 subjects from the ADNI dataset demonstrate superior performance with 92.3% overall accuracy for three-class classification (NC/MCI/AD). Binary classification tasks achieve strong performance: AD vs. NC (94.7% accuracy, AUC: 0.967), MCI vs. NC (85.7% accuracy, AUC: 0.918), and AD vs. MCI (88.3% accuracy, AUC: 0.942), outperforming state-of-the-art methods. Furthermore, external validation on an independent OASIS-3 cohort confirms the model's robust generalization capability. Ablation studies validate that adding HAFP to the baseline improves accuracy by 20.7 percentage points, and further incorporating TA yields an additional 4.5 percentage-point gain. The directional attention mechanism represents a novel approach for exploiting anatomical plane-specific information in 3D neuroimaging analysis. The current implementation remains limited to single-time-point sMRI and requires longitudinal, multimodal, prospective, and lightweight-deployment validation before routine clinical use.