Alzheimer's Disease Classification Using CUDA-Accelerated DenseNet201 and VGG16

Automated classification of Alzheimer's disease through MRI analysis presents significant computational challenges due to complex deep learning architectures and largescale image processing requirements. This paper introduces a GPU-accelerated multi-branch neural network framework that combines modified DenseNet201 and VGG16 architectures for efficient Alzheimer's disease stage classification. The proposed model employs parallel feature extraction pathways, with each branch independently processing MRI inputs before feature fusion. CUDA optimization techniques are implemented across the processing pipeline, including data loading, augmentation, and model training phases, resulting in significantly reduced computational overhead. The framework classifies brain MRI scans into four categories: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. Experimental results demonstrate that the proposed hybrid architecture achieves 98% classification accuracy while substantially reducing training time through GPU acceleration. The model's performance is further enhanced through adaptive learning rate scheduling and selective layer fine-tuning. This work establishes an efficient approach for deploying complex neural architectures in clinical neuroimaging applications where computational resources and processing time are critical constraints.