An Interpretable Dual-Path Convolutional Network for Multi-Class Alzheimer’s Disease Staging from Structural MRI

Abstract

Alzheimer disease (AD) is a developmental neurodegenerative disorder impacting more than 55 millionindividuals worldwide and early, accurate staging is an elusive clinical necessity. This paper introduces a new deep learningsystem that is capable of staging Alzheimer disease in multiple classes automatically using structural Magnetic ResonanceImaging (MRI). Our architecture (Dual-Path Convolutional Neural Network (DP-CNN)) constitutes parallel ResNet-50 andVGG-16 feature extractors with a spatial squeeze-and-excitation attention component and a dual-path fusion block. Themodel was trained and tested on a pool of 6,400 T1-weighted MRI scans (combined) on two groups of participants: OASIS-3 (n=1,378 participants) and ADNI (n=3,900 scans) representing four diagnostic groups: Cognitively Normal (CN), MildCognitive Impairment (MCI), Early Alzheimer’s Dise The proposed model had a total classification accuracy of 97.4, amacro-averaged sensitivity of 96.8, specificity of 97.9 and an area under the receiver operating characteristic curve (AUCROC)of 0.993 in the held-out test set. Experiments on ablation proved independent contributions of both the dual-pathfusion ( +4.3% accuracy over the best single-backbone baseline) and spatial attention module ( +1.8%). Generalizationtesting (trained on ADNI, tested on OASIS-3) had an accuracy of 94.2% meaning the results are robust across scannerenvironments. The gradient-weighted Class Activation Mapping (Grad-CAM) has shown that model saliency is concentratedin the hippocampus, parahippocampal gyrus, and anterior cingulate cortex, which are considered to be canonical in ADneuropathology, and gave neuroimaging-grounded interpretability. These findings make the suggested DP-CNN a scalable,non-invasive, and clinically interpretable device to detect early Alzheimer disease.