Dense and accurate automated dementia stage classification from MRI scans is still a challenging task because of the inconsistencies in structural pat- terns and the inability of current deep learning models to manage both local and global feature representations effectively. In this work, a new hybrid ensemble model is proposed that combines CNN-based (ResNet-50, ConvNeXt, DenseNet-121) and transformer-based (Swin Transformer, ViT) models to synergistically extract their complementary strengths for effective dementia classification. An explicitly crafted preprocessing pipeline, including N4 bias field correction, CLAHE, non-local means denoising, Sobel edge detection, and K-means segmentation, provides high-quality input for enhanced feature extraction. The model was tested on a dataset of 87,000 MRI scans with 8500 images independently tested. Experiment results show Swin Transformer obtained AUC-ROC scores of 0.99 for Mild Dementia and 0.97 for Moderate Dementia, and the ensemble model delivered balanced decisions for all stages of dementia. Cohen’s Kappa (0.57) and Matthews Correlation Coefficient (0.60) established model reliability, and class-wise precision-recall substantiated its discrimination capacity. Furthermore, Grad-CAM and attention maps enabled model interpretability by emphasizing primary regions driving predictions, thereby improving clinical relevance. The results highlight the effectiveness of transformer-CNN ensembles in neuroimaging-based dementia diagnosis and promise their suitability for use in real-world diagnostic assistance.