Explainable Hybrid Deep Learning Framework for Alzheimer’s Disease Classification using Cross-Attention Fusion

• Novel Attention-Guided Hybrid Framework:In this
paper, we propose highly capable and attention-based
deep learning framework, that combines ConvNeXtV2-
Tiny and EfficientNetV2-S using Cross-Attention Fusion
and Convolutional Block Attention Modules (CBAM) to
achieve accurate classification of the four Alzheimer’s
disease classes.
• High Classification Performance: The proposed frame-
work leads to an excellent overall test accuracy of 97.33%
and a Macro ROC-AUC of 0.9996 indicating highly
robust and reliable performance for automated early de-
mentia detection.
• Explainable AI for Model Interpretation: To improve
model interpretability, we integrate a comprehensive XAI
framework incorporating ten CAM variants, including
Grad-CAM, together with SHAP and LIME to provide
detailed visual explanations of the model’s predictions.
• Deep Latent Space and Misclassification Analysis:
We perform a thorough analysis of boundary cases us-
ing a purpose-built misclassification analysis and t-SNE
and UMAP clustering. These visualizations convert raw
model outputs into clinically meaningful insights about
structural decision boundaries.