NeuroStageNet: An Explainable Deep Learning Framework for Accurate Multi-Stage Senile Dementia Detection from Brain MRI Using CNNs and Grad-CAM

Abstract

Senile Dementia disease (SDD) is a progressive neurodegenerative condition currently affecting over 55 million people worldwide, and timely detection at early stages remains a persistent clinical challenge. While deep learning models have shown strong diagnostic accuracy on brain MRI data, most operate as black-box systems - clinicians cannot inspect the model's reasoning, which undermines trust and clinical adoption. This paper presents a CNN-based multi-class classification framework augmented with Gradient-weighted Class Activation Mapping (Grad-CAM) to simultaneously predict AD severity and generate spatially interpretable saliency maps on MRI slices. Trained and evaluated on the OASIS-3 neuroimaging dataset comprising 1,098 axial MRI scans across four severity classes, the proposed architecture achieves an overall classification accuracy of 94.7%, outperforming VGG-16 (89.2%), ResNet-50 (91.4%), InceptionV3 (90.1%), and AlexNet (86.3%). Grad-CAM overlays consistently highlight the hippocampal and entorhinal cortex regions, aligning with established neurological markers of AD progression. The dual output of high accuracy and regionlevel transparency positions the proposed system as a practical and trustworthy tool for clinical decision support in early AD screening.