Skip to content
Conference

A Contrast-Attention ResNet50 for Robust Cross-Dataset Alzheimer's Disease Classification from Structural MRI

Aug 2026 · International Conference Computational Vision and Bio Inspired Computing · pp. 1046-1053 · 0 citations · 19 references

Abstract

Alzheimer's disease (AD) is a progressive neurodegenerative condition where early detection can delay the disease progression and improve the quality of life of the people concerned. Features like hippocampal atrophy, cortical thinning, and ventricular enlargement are captured by structural MRI for AI-assisted diagnosis, single-dataset, and lack cross-dataset validation still remain a exciting approach. This paper reviews major AD-MRI benchmarks (OASIS, Kaggle, Zenodo) and proposes a ResNet-50 enhanced with a Contrast Attention Module (CAM) to improve subtle structural differences in early-stage impairment across four classes: No, Very Mild, Mild, and Moderate Impairment. The model achieved 92% precision for Mild Impairment (Train: Kaggle+Zenodo, Test: OASIS) and 94% precision for Moderate Impairment(Train: Kaggle+OASIS, Test: Zenodo).OASIS emerges as the accurate training source (0.9356 accuracy while testing on Kaggle), while Zenodo generalizes poorly (when tested on OASIS as low as 0.65). Interpretability analyses using Grad-CAM, channel attention, and t-SNE indicate that the model consistently focuses on periventricular regions. However, errors in Very Mild cases suggest that it still depends heavily on ventricular size and struggles to capture the subtle hippocampal changes that distinguish early dementia from normal aging.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.