MRI-Based Alzheimer’s Disease Classification Using ResNeXt50 with Label Smoothing

Abstract

Alzheimer's disease (AD) is a neurodegenerative disorder characterized by progressive cognitive decline. Identifying the condition at an early stage is essential for timely intervention and diagnosis. To improve MRI-based four-class AD classification, this study employs ResNeXt50 with Label Smoothing (LS) for four-class AD recognition. Public Magnetic Resonance Imaging (MRI) images were preprocessed and augmented during training using random horizontal flipping, slight rotation, translation, and scaling. The images were then classified by ResNeXt50, where LS was added to the cross-entropy (CE) loss to make it more generalizable. ResNeXt50 uses ResNeXt Blocks with grouped convolution to learn multi-branch features while controlling model complexity. The accuracy rate of this model is 99.50%, while the precision rate, recall rate and F1 score all exceed 99.5%, and the Kappa coefficient is 0.9936. The ablation experiments proved that LS could enhance performance, but other measures such as class weighting and extra attention modules did not produce any further improvements. The research results indicate that ResNeXt50 with LS can effectively enhance the AD classification performance of MRI images. It offers a simple and effective solution for intelligent-assisted diagnosis of AD.