Optimizing MRI-Based Alzheimer’s Disease Detection Using EfficientNet and Explainable Artificial Intelligence
Abstract
Alzheimer’s disease has been difficult to diagnose early because structural changes in the brain often appear subtle and overlap with normal aging. This study aimed to improve the accuracy and interpretability of Alzheimer’s disease detection from structural magnetic resonance imaging by developing a classification framework based on the EfficientNet B3 architecture. The imaging data were processed through grayscale conversion, resizing, contrast enhancement, and normalization, followed by synthetic oversampling to balance the diagnostic categories. The model demonstrated strong performance, achieving an accuracy of 0.9965, high precision and recall across all classes, perfect discrimination scores, and a high Matthews correlation coefficient. Gradient-weighted Class Activation Mapping and Shapley-based explanations revealed that the model focused on neuroanatomical regions consistent with known markers of Alzheimer’s disease. The results showed that the proposed framework met the study objective and produced reliable and clinically relevant predictions. These findings also suggested opportunities for future expansion through multimodal imaging and integration into clinical decision-support systems.
