SHAP-Based Explainability for Local and Global Insights in Alzheimer's Detection

Alzheimer's disease is a progressive neurodegenerative disorder that leads to cognitive decline and loss of independence, making early and accurate diagnosis essential. Recent advances in Machine Learning (ML) have enhanced medical image analysis, but the opaque nature of deep learning models limits their adoption in clinical practice. This study introduces SCR NetX, a CNN model based on the VGG 16 architecture, to classify Alzheimer's disease into four stages: non demented, very mild, mild, and moderate dementia. To improve interpretability, the model integrates Explainable AI (XAI) using SHAP (SHapley Additive eXplanations) for both local and global analyses. Local explanations highlight MRI regions that influence individual predictions, aiding in case-specific evaluation, while global explanations reveal the overall behavior of the model. Two segmentation methods—grid-based for broad region analysis and SLIC (Simple Linear Iterative Clustering) for fine-grained superpixel analysis—are employed to ensure precise and clinically interpretable outputs. This framework combines accurate classification with transparent decision-making, bridging the gap between AI-driven diagnostics and practical clinical application.