Explainable Federated Attention-Based Deep Learning for Alzheimer’s Disease Detection

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder in which early and reliable detection is critical for clinical management. This study proposes a privacy-preserving and explainable deep learning framework for Magnetic Resonance Imaging (MRI) -based AD classification that targets predictive performance, data confidentiality, and interpretability. The framework is based on an EfficientNetB0 backbone enhanced with a Convolutional Block Attention Module (CBAM). To address institutional data-sharing constraints, the model was trained in a federated learning setting, where local model updates were aggregated using Federated Averaging without sharing raw MRI data. Model interpretability was supported using Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations. The classification task was formulated as a binary distinction between cognitively normal subjects and subjects with Alzheimer’s-related impairment. Experimental evaluation on a publicly available MRI dataset showed federated performance close to centralized training, with 94.66% final test accuracy compared with 95.70% under centralized training. On an external MRI validation dataset, the model achieved 95.00% accuracy, supporting generalization to unseen data distributions. Grad-CAM heatmaps indicated that the model focused on brain regions rather than background artifacts. These results suggest that federated and explainable attention-based deep learning can support collaborative, privacy-aware AD screening while maintaining competitive predictive performance.