A Comparative Study of CNN and Vision Transformer Models in Alzheimer's Detection

In this work, a novel deep learning framework the Hierarchical Attention Network (HierAttNet) is proposed to replace the previously used Vision Transformer (ViT-B/16) and all CNN baselines in the literature to stage Alzheimer's disease (AD) from MRI brain images. The ConvNeXt-Small feature extractor and the two-scale cross-attention module alternately attend to the local patch level and the region level in a hierarchical fashion, which is not reported in the AD neuroimaging literature. We compare HierAttNet with six additional state-of-the-art algorithms, which were not presented in the original paper: EfficientNetV2-S, Swin-T, MobileViT-S, CrossViT-15, DeiT-S and ConvNeXt-S. The following experiments are performed on the Kaggle Alzheimer's MRI dataset (6,400+ images, 4 stages of Alzheimer's). HierAttNet obtains 99.91% accuracy and F1=0.9993 outperforming all the evaluated models and even the previous state-of-the-art ViT-B/16 (99.87%). The performance of all models consistently dropped with data augmentation. In this work, we introduce HierAttNet as a new state-of-the-art in Alzheimer's MRI classification, and a stringent benchmark for subsequent multimodal diagnostics studies.