A Hybrid CNN Ensemble Approach for Intelligent Alzheimer’s Disease Detection in MRI Imaging

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative condition, and diagnosing it early and accurately is critical for better clinical care. This study introduces a deep learning framework for multiclass AD detection using brain MRI images. The model classifies patients into four groups: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. The approach combines standardized MRI preprocessing, oversampling of the training set to address class imbalance, and transfer learning based on an EfficientNetV2S backbone with a custom classification head. Using a dataset of 6,400 MRI images, the model was trained in two stages to strengthen task-specific feature learning.When tested on a separate set of 640 images, the framework reached 93% accuracy, with a macro precision of 0.80, macro recall of 0.85, and macro F1-score of 0.81. AUC values were above 0.90 across all classes. The model showed strong sensitivity in identifying early and intermediate stages of the disease. Most errors appeared between neighboring classes, which aligns with the gradual progression of Alzheimer’s. These results suggest that the framework is effective at detecting clinically meaningful MRI patterns and has strong potential as an AI-assisted decision-support system for Alzheimer’s diagnosis. The study still has a few limitations. Evaluation was carried out on only one dataset, the Moderate Demented class had a limited number of samples, and external validation was not included. Even with these constraints, the findings show that deep learning-based MRI analysis holds strong promise for making Alzheimer’s detection more reliable and scalable.