Early Detection of Alzheimer's Disease: A Deep Learning Approach for Accurate Diagnosis

The research delineates the development and evaluation of a deep learning paradigm for the stratification of Alzheimer's disease phases, leveraging neuroimaging and transfer learning methodologies. The proposed architecture encompasses a suite of models, including DenseNet, GoogLeNet, ResNet18, EfficientNet, and RegNet, aiming to ascertain the most effective model for amalgamation into a novel framework, denominated as AlzheimerTriMatterNet. This framework undertakes the analysis of discrete cerebral compartments—namely white matter, gray matter, and the unaltered brain—utilizing explainable artificial intelligence (XAI) mechanisms, with a particular emphasis on Gradient-Weighted Class Activation Mapping (Grad-CAM) to underscore pivotal diagnostic regions. Empirical analysis is conducted utilizing the Alzheimer MRI Preprocessed Dataset sourced from the University of Northern British Columbia (UNBC) and Kaggle, with the evaluation of model efficacy predicated on accuracy metrics and confusion matrices. The empirical outcomes manifest the superior efficacy of the ResNet18 model, which exhibited an accuracy of 82.5%, whereas the integrated AlzheimerTriMatterNet model, incorporating ResNet18, registered an accuracy of 88.4%.