Attention-Based RNN for Alzheimer's Diagnosis: A Multi-Modal Approach with Enhanced Feature Retention Algorithm

Alzheimer's disease (AD) affects millions, and early detection is key. This paper introduces a new deep learning algorithm, Attention-Based Recurrent Neural Network (ABRNN), to detect AD using multi-modal data, including MRI scans and clinical features. ACO is used for preprocessing to enhance image quality and retain important patterns in MRI scans for diagnosis. SegNet isolates the brain regions that indicate the progression of AD through image segmentation. ResNet is used for feature extraction to capture deep hierarchical features from the segmented MRI scans to identify subtle but important features of brain degeneration. A memory-augmented block is added to store key feature patterns so the network can recall important information for AD diagnosis. A hybrid attention mechanism combines imaging and non-imaging features to focus on the relevant brain regions and clinical indicators. ABRNN is compared to 4 existing algorithms: CNN, LSTM, VGG-16, and Random Forest. In experiments, ABRNN achieved 97.85% accuracy, outperforming CNN (94.2%), LSTM (93.5%), VGG-16 (95.1%), and Random Forest (92.3%). ABRNN's performance is better because it retains important features through memory augmentation, extract features well with ResNet, and segments well with SegNet, so it's more accurate than traditional methods