An interpretable ensemble CNN framework for MRI-based diagnosis of neurodegenerative diseases

Abstract

In medical imaging, convolutional neural networks have made major advances, particularly in the classification of neurodegenerative disorders like Alzheimer's and Parkinson's disease. But their 'black-box' nature makes model interpretability difficult, which is crucial in healthcare for confidence and informed decision-making. This research introduces a multiclass multiclass transfer learning-based ensemble classification system that classifies brain MRI images into AD, PD, and healthy controls using EfficientNetB7, ResNetV2-50, and DenseNet121. Soft voting determines categorisation by averaging predictions from all three models and selecting the class with the highest likelihood. Grad-CAM (Gradient-weighted Class Activation Mapping) visualises brain areas impacting model predictions to improve interpretability. This method increases classification performance and provides transparent insight into decision making, increasing trust in AI-assisted neuroimaging applications.