Cross- Architectural Evaluation of AI Models for Dementia Stage Classification

Alzheimer's disease is a progressive neurodegenerative disorder characterized by the deterioration of memory and higher-order cognitive functions, underscoring the importance of its early detection. The present study investigates the automatic identification and classification of Alzheimer's disease from MRI scans by leveraging advanced deep-learning techniques. A dataset of 12,020 MRI images was employed, partitioned into four diagnostic categories: mild, moderate, non-demented, and very mild. To ensure both accuracy and computational efficiency, rigorous preprocessing was applied-images were uniformly resized and normalized, while Contrast-Limited Adaptive Histogram Equalization (CLAHE) and a suite of geometric and photometric augmentations were used to enhance feature visibility and bolster model generalization. Multiple convolutional neural network (CNN) architectures were evaluated, including ResNet-50, VGG-16, VGG-19, DenseNet-201, Inception-V3, MobileNet-V2, Xception, and EfficientNet-V2. In addition, hybrid ensemble frameworks were constructed by fusing ResNet-50 with VGG-16, VGG-19, Inception-V3, and MobileNet-V2 backbones, and integrating these with machine-learning classifiers such as XG-Boost, k-Nearest Neighbors, and Random Forests. By combining deep, hybrid, and ensemble approaches, this work aims to detect subtle patterns across MRI sequences and thereby enhance the predictive accuracy of Alzheimer's disease classification.