An explainable hybrid CNN–vision transformer framework for accurate multistage Alzheimer’s disease classification using brain MRI

Early and reliable diagnosis of Alzheimer’s disease (AD) is currently one of the biggest challenges in neuroscience-related image analysis due to the subtle intensity variations in the brain structure. The classical deep models work quite effectively in diagnosing AD; however, the results are infeasible. In this context, this research work proposes a combination approach containing CNN models along with Vision Transformers (ViTs) to recognize the multiclass classifications in Alzheimer’s diseases from brain MRI images utilizing the principles of XAI. The proposed approach was implemented on the Alzheimer’s Disease Multiclass Dataset containing 44,000 skull-stripped MRI images belonging to the categories of Non-Demented, Very Mild Demented Patients, Mild Demented Patients, and Moderate Demented Patients. The proposed approach also applies data augmentation techniques along with transfer techniques to cope with the problem of overfitting. The proposed approach also applies the principles of visualization techniques like Grad-CAM, SHAP values, and LIME. The hybrid CNN-ViT model has achieved an overall accuracy of 99.18%, achieving the highest performance among the evaluated baseline models. The obtained improvements were verified as highly significant using statistical analysis. The Grad-CAM visualizations showed good agreement between model activation patterns and known biomarkers, notably for the hippocampus/mesial temporal lobe, which was verified using the Explainability Concordance Index of 0.91. The proposed research has shown that integrated CNN-based models answering transformer-based architectures defined using an interpretable deep learning paradigm can provide a very accurate, showing neuroanatomical spatial concordance, and meaningful model for the detection of AD. The proposed study opens the door for the development of interpretable AI systems systems towards providing a clear interpretation for clinical AI tools for the diagnosis of AD.