Motivation: The CNN-ViT model, equipped with CNN and vision transformer (ViT) branches, can capture and then fuse local features and global representations of images, potentially improving the performance of brain-age prediction. Goal(s): To assess the performance of brain-age prediction model in CNN-ViT architecture. Approach: The 3D-CNN-ViT brain-age model was trained with T1 weighted images, followed by two comparative experiments and a clinical experiment to assess its performance superiority. Results: The CNN-ViT brain-age model outperformed both the CNN and ViT one. Besides the whole brain, this architecture also effectively predicted brain-age in individual lobes, where the brain-aging acceleration was more sensitive to dementia. Impact: The fused local and global features of MRI data improve the performance in brain-age prediction paradigm, suggesting that the CNN-ViT architecture has potential to promote prognosis prediction or biotype classification in clinical applications using MRI data.
Gan et al. (Tue,) studied this question.