Text-Enhanced Vision-Language Motion Correction (VLM-MoCo) for Mitigating Severe Motion Artifacts in MRI Scans

Motivation: MRI images often suffer from motion artifacts due to patient movement, compromising image quality and leading to diagnostic inaccuracies. Goal(s): Enhance artifact removal by integrating textual descriptions into deep learning methods. Approach: We developed the Vision-Language Motion Correction (VLM-MoCo) method, combining image data with textual descriptions of artifact characteristics. This approach leverages a BERT-encoded framework integrated into a 3D pix2pix GAN. Results: VLM-MoCo significantly outperformed the baseline, achieving lower NMSE and higher PSNR and SSIM values, demonstrating its effectiveness in improving image quality and artifact removal. Impact: By integrating text descriptions into deep learning models, this method significantly enhances collaboration between clinicians and AI systems to remove MRI motion artifacts. It especially benefits patients prone to involuntary movements and transforms clinician-AI collaboration in medical imaging.