Deep learning feature fusion with transfer learning for Alzheimer’s disease classification

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by irreversible cognitive decline and memory impairment, for which early and accurate diagnosis is critical to enabling timely clinical intervention and disease management. Deep learning-based analysis of structural magnetic resonance imaging (MRI) has emerged as a promising approach for automated dementia staging, offering the ability to detect subtle neuroanatomical changes associated with disease progression that may not be easily identified through conventional clinical assessment. This study investigates transfer learning and multimodel feature fusion for four-class Alzheimer’s disease severity classification using 6,400 T1 weighted MRI images obtained from a publicly available Kaggle dataset. The dataset includes four categories: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. Three pretrained convolutional neural network architectures VGG16, ResNet50, and MobileNetV3 were independently fine tuned using ImageNet initialized weights. To evaluate the effectiveness of feature-level integration, four fusion strategies feature concatenation, element-wise addition, element-wise multiplication, and attention-based fusion were implemented across three backbone pairings: VGG16-ResNet50, ResNet50-MobileNetV3, and VGG16-MobileNetV3. All models were trained for 60 epochs using the Adam optimizer and evaluated using accuracy, precision, recall, and F1-score. Among the standalone models, VGG16 achieved the highest performance, with a test accuracy of 99.48%, precision of 0.9929, recall of 0.9927, and F1-score of 0.9927. Among fusion configurations, VGG16-ResNet50 concatenation achieved the best result with 96.25% accuracy. Attention-based fusion exhibited mixed performance across backbone pairings. While the ResNet50 and MobileNetV3 attention model yielded the lowest accuracy of 82.08%, the VGG16 and MobileNetV3 attention model achieved a strong accuracy of 97.40%, representing the best-performing attention configuration evaluated in this study. To further assess generalizability and address the limitations of image-level splitting in the original dataset, the three best-performing configurations were additionally evaluated on an independent OASIS-derived MRI dataset using patient-wise partitioning. Because the Moderate Dementia category contained only two unique subjects, the external validation was performed as a three-class classification task comprising Non-Demented, Very Mild Dementia, and Mild Dementia, with no patient overlap across the training, validation, and testing subsets. Notably, although no fusion strategy surpassed standalone VGG16, the VGG16 and MobileNetV3 attention model achieved competitive performance with 97.40% accuracy, indicating that a well-optimized single backbone can outperform multimodel feature fusion under certain dataset conditions. These findings demonstrate the effectiveness of transfer learning for MRI-based Alzheimer’s classification and emphasize the need for careful design and sufficient data when applying feature-level fusion strategies.