Hybrid 3D CNN and ResNet Deep Transfer Learning for High-Resolution Hippocampal Atrophy Mapping and Automated Alzheimer’s MRI Diagnosis

Early and accurate detection of Alzheimer's disease (AD) is crucial for a timely clinical intervention. Atrophy of the hippocampus has been established as a key neurodegenerative biomarker. This study presents a Hybrid 3D CNN–ResNet model for automated hippocampal segmentation and dementia classification using high-resolution MRI scans. The proposed framework integrates 3D U-Net-based hippocampal segmentation with multi-scale feature extraction and deep residual learning, enabling a precise atrophy quantification and robust classification across the AD stages. A standardized preprocessing pipeline, incorporating NIfTI conversion, spatial normalization, denoising, and contrast enhancement, ensures consistency across multi-site datasets. The model was optimized using AdamW, Cyclical Learning Rate (CLR), and early stopping, achieving 97.31% classification accuracy and a 92.84% Dice Similarity Coefficient (DSC) for hippocampal segmentation. Grad-CAM and SHAP-based interpretability confirm the biologically meaningful feature representations, aligning with established hippocampal atrophy patterns in AD progression. The external validation on the OASIS dataset demonstrated strong generalization, with only a 2.3% accuracy reduction, underscoring the model’s robustness and clinical applicability. These findings establish the proposed approach as an effective and interpretable deep learning framework for early AD diagnosis and longitudinal disease monitoring.