<![CDATA[Early, accurate diagnosis of Alzheimer’s Disease (AD) is vital for effective intervention. Properly classifying its progression from cognitively normal to moderate dementia is essential for tailoring treatment and management plans. The proposed research is a hybrid deep learning framework that integrates the EfficientNet-B3 and the ResNet50 with sophisticated attention units in order to classify the MRI scans of Alzheimer's disease in multi-classes. The model proposed combines both a Convolutional Block Attention Module (CBAM) based on the refinement of channels and space features and Multi-Head Self-Attention based on cross-branch feature interaction. The dual-branch architecture yields complementary features, with EfficientNet-B3 being able to pick fine-grained patterns and ResNet50 being able to pick strong hierarchical representations. The characteristics in both branches are mapped to 512 dimensions, operated by multi-head attention classification. The model and extensive preprocessing were implanted on a series of 33,984 augmented Alzheimer's MRI images in four categories (MildDemented, ModerateDemented, NonDemented and VeryMildDemented. This hybrid model had outstanding performance of 98.21% test accuracy, 98.23% precision, 98.21% recall, and 98.21% F1-score, which was far much better than the accuracies of other baseline architectures such as VGG16 (74.11%), ResNet50 (93.68%), EfficientNetB0 (63.52%), DenseNet121 (64.12%), and CustomCNN (68.87%). These findings support the usefulness of hybrid systems consisting of attention mechanisms to diagnose Alzheimer's disease automatically by using neuroimaging information.]]>
