HCT-EA: A Hybrid CNN–Transformer with Explainable Attention for Early Alzheimer's Detection

Structural magnetic resonance imaging (sMRI) has become a valuable biomarker for identifying neuroanatomical alterations associated with Alzheimer’s disease (AD); however, accurately recognizing the subtle morphological changes that characterize early disease progression remains a significant challenge. Existing deep learning models exhibit complementary strengths and limitations: convolutional neural networks (CNNs) effectively preserve fine-grained local anatomical information but inadequately model long-range relationships across distributed brain regions, whereas Vision Transformers (ViTs) capture global contextual dependencies through self-attention yet often fail to retain detailed structural abnormalities and typically require large-scale training datasets. To overcome these limitations, this paper proposes a Hybrid CNN–Transformer with Explainable Attention (HCT-EA), a dual-branch architecture that simultaneously learns local neuroanatomical features and global contextual representations from 3D sMRI volumes. The proposed framework comprises (i) a CNN-based feature extraction backbone for fine-grained structural representation, (ii) a transformer encoder for modeling long-range interregional dependencies, (iii) a bidirectional cross-attention fusion module that adaptively integrates local and global features, and (iv) an intrinsic explainable attention mechanism that generates clinically interpretable heatmaps highlighting disease-relevant brain regions. Experimental evaluation on the ADNI dataset demonstrates that HCT-EA achieves a 98.64% classification accuracy for AD versus healthy control (HC), surpassing BMSSnet (97.35%), DuoViT-AD (98.56%), and conventional ViT models (94.21%). For early prediction of MCI converters versus non-converters, the proposed model attains 89.12% accuracy, representing improvements of 3.98% over BMSSnet and 6.58% over standard ViT. Furthermore, HCT-EA achieves sensitivities of 97.76% for AD detection and 88.45% for MCI conversion, demonstrating strong capability for early-stage diagnosis. Ablation studies reveal that the cross-attention fusion module contributes a 2.3% improvement in classification accuracy, while the explainable attention mechanism produces saliency maps that correlate significantly with hippocampal volume (r = 0.79, p < 0.001) and entorhinal cortex thickness (r = 0.74, p < 0.001), supporting the clinical relevance of the learned representations. These findings indicate that HCT-EA provides an effective and interpretable framework for early Alzheimer’s diagnosis by combining complementary local and global feature learning with adaptive attention-based fusion and computational efficiency suitable for clinical deployment.