HYBRID TRANSFORMER-ENHANCED 3D CNN FOR EXPLAINABLE ALZHEIMER’S DISEASE CLASSIFICATION AND PROGRESSION PREDICTION

Abstract

Alzheimer’s disease (AD) is an irreversible neurodegenerative disorder whose timely and interpretable diagnosis remains a major clinical challenge. Three-dimensional convolutional neural networks (3D CNNs) capture fine-grained local morphometric changes from structural magnetic resonance imaging (sMRI), yet their limited receptive field hinders modelling of long-range anatomical dependencies, while pure transformer models are data-hungry and difficult to interpret. This paper proposes a hybrid Transformer-Enhanced 3D CNN (TE-3DCNN) that couples a squeeze-and-excitation residual 3D CNN encoder with a transformer encoder through a gated cross-module fusion mechanism, and augments it with a dedicated progression-prediction head for mild cognitive impairment (MCI)-to-AD conversion. To make predictions trustworthy, the framework integrates Grad-CAM++ and self-attention rollout to generate volumetric saliency maps. Experiments on the public ADNI sMRI cohort for three-way classification (cognitively normal, MCI, AD) show that TE-3DCNN attains 94.8% accuracy, 94.7% macro F1-score and 0.979 mean AUC, outperforming a 3D ResNet, a 3D Vision Transformer and a convolution–Swin transformer baseline by 2.7–6.4 percentage points. Ablation studies confirm the complementary contribution of the CNN encoder, transformer branch, gated fusion and squeeze-and-excitation recalibration. The produced saliency maps consistently localise the hippocampus, medial temporal lobe and ventricular regions, agreeing with established AD neuropathology. The results indicate that TE-3DCNN offers an accurate, explainable and clinically meaningful tool for AD diagnosis and prognosis.