An Explainable Hybrid Deep Learning Framework for Multi-Class Alzheimer's Disease Classification from Brain MRI Using MobileNetV3–ResNet50 Feature Fusion

Abstract

Early diagnosis of Alzheimer disease (AD) is a progressive neurodegenerative disorder is key to initiating timely therapeutic interventions. This paper proposes a deep learning model for four-class AD classification using brain MRI. We use a two pre-trained networks (MobileNetV3 and ResNet50) fusion with a final classifier of Multi-Layer Perceptron (MLP). The model obtained F1 scores of 0.88, 0.85, 0.87, and 0.93 for the NonDemented, VeryMildDemented, MildDemented and ModerateDemented classes respectively and had a perfect recall rate of 100% for the ModerateDemented class. Grad-CAM visualizations which represented the framework's ability to provide high prediction confidence as well as a pathologically relevant brain region identification, were able to correctly predict the class with 98.6% accuracy.