Explainable AI for Dementia Classification: Understanding Deep Learning Predictions

Abstract

Deep learning has revolutionized medical image classification, yet challenges remain in interpretability and model generalization. This study proposes an explainable AI framework for dementia classification using five deep learning models CNN, VGG16, ResNet50, MobileNet, and EfficientNet—evaluating their performance on a large-scale brain scan dataset (9,488 images) across four dementia stages (Mild, Moderate, Very Mild, and Non-Demented). Among them, EfficientNet achieved the highest accuracy (92.1%), benefiting from compound scaling for optimized feature extraction. Additionally, explainability techniques (Grad-CAM and SHAP) were employed to visualize model decision-making, addressing the critical need for interpretable AI in healthcare. The study also explores dataset bias and class imbalance, implementing data augmentation and weighted loss functions to enhance fairness. Comparative analysis with state-of-the-art models (ViTs, DenseNet, and Swin Transformer) provides a benchmark for future research. The findings highlight the importance of model scalability, explainability, and computational efficiency in real-world deployment.