Abstract
Background: Motion during magnetic resonance imaging (MRI) scans is common and can degrade image quality and reduce diagnostic accuracy. This study developed MC-Former, a motion correction method for brain MRI based on vision transformers and self-supervised learning. Methods: Motion-free brain MRI data from 52 subjects were used to generate 13,700 training, 1,950 validation, and 4,680 test axial slices from 3D sagittal T1-weighted scans acquired at 3T. MC-Former was derived from a vision transformer-based encoder-decoder architecture with skip connections. During self-supervised pre-training, 50% of patches from motion-free images were randomly masked and reconstructed; synthetic motion-corrupted images were then used for supervised training. Evaluation was performed on simulated motion-corrupted axial images unseen during training, using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). Results: MC-Former improved PSNR from 26.4 to 32.8 and SSIM from 0.77 to 0.94, significantly outperforming MC-Net and Restormer on the test set. For motion-free images, MC-Former achieved SSIM > 0.99. Conclusion: These results indicate that MC-Former effectively suppresses simulated motion artifacts in brain MRI while preserving artifact-free image quality. The code and trained model are available at https://github.com/MRIMoCo/MC-Former.