Classification of Alzheimer's Disease via Vision Transformer

Deep models are powerful in capturing the complex and non-linear relationship buried in brain imaging data. However, the huge number of parameters in deep models can easily overfit given limited imaging data samples. In this work, we proposed a cross-domain transfer learning method to solve the insufficient data problem in brain imaging domain by leveraging the knowledge learned in natural image domain. Specifically, we employed ViT as the backbone and firstly pretrained it using ImageNet-21K dataset and then transferred to the brain imaging dataset. A slice-wise convolution embedding method was developed to improve the standard patch operation in vanilla ViT. Our method was evaluated based on AD/CN classification task. We also conducted extensive experiments to compare the transfer performance with different transfer strategies, models, and sample size. The results suggest that the proposed method can effectively transfer the knowledge learned in natural image domain to brain imaging area and may provide a promising way to take advantages of the pretrained model in data-intensive applications. Moreover, the proposed cross-domain transfer learning method can obtain comparable classification performance compared to most recent studies.