Abstract

Alzheimer's Disease (AD) classification from brain MRI images remains a strenuous mission due to the intricacy of the disease and limited dataset sizes. Eventhough Convolutional Neural Networks (CNNs) have excelled in the classification of brain diseases using MRI data, they are incompetent to apprehend global dependencies. Also, their results are not interpretable which is a major problem in medical domain. Transformer uses attention mechanisms to go with or even surpass CNNs on various vision tasks. This study proposes a novel fusion model integrating the complementary advantages of DenseNet-121 and Vision Transformer to address these challenges. By synergizing the strengths of both architectures, the proposed fusion model extracts comprehensive image features. To further optimize feature discrimination and computational efficiency, an ExtraTree classifier-based feature selection technique is incorporated. The performance of the proposed model is evaluated using standard metrics and compared with state-of-the-art techniques. Results demonstrate a superior classification accuracy of 99%, with the fusion model effectively differentiating between various AD stages. Furthermore, Class Activation Maps (CAMs) are utilized to visualize the model's decision-making process thereby enhancing trust in the model's predictions. We also provided a visual comparison of Grad-CAM, Grad-CAM++, Score-CAM, and Faster Score-CAM visualization techniques to assess the performance of these in highlighting discriminative regions for AD classification.