This paper describes a comprehensive framework for segmentation, classification, and 3D reconstruction of brain MRI scans for structural change analysis in dementia. It combines the classification with EfficientNet, the segmentation with U-Net, and 3D reconstruction with the Marching Cubes algorithm. Data uniformity was ensured and contrast enhanced by applying techniques like normalization, denoising, and Otsu thresholding before the actual analysis. EfficientNet obtained classification accuracy of 98.6% in distinguishing six categories of dementia, while the U-Net achieved the segmentation accuracies of 88.86 and 90.88% for gray and white matters, respectively. The high-resolution 3D visualizations were facilitated by the Marching Cubes algorithm. Significant morphological differences along the stages of dementia were discovered. The volume measurements and the Hausdorff distance, ranging from 20.91 to 42.87, quantified changes in structure, thus better facilitating an understanding of disease progression. It thereby shows potential for 2D-3D deep learning model combinations, in both accurate diagnosis and monitoring of dementia. In further work, it seeks to make the diversity of the datasets larger while improving model generalization by incorporating real-time diagnostic abilities into clinical practices, as a transformation tool for managing neurodegenerative disease.