A Physics-Informed Deep Learning Method for Correcting Motion Artifacts in Brain MR Imaging

Motivation: Motion artifacts frequently degrade MRI quality, necessitating costly repeat scans and increasing patient discomfort. Current deep learning methods often induce hallucinations, especially at high levels of artifact presence. Goal(s): We aimed to develop a more robust, physics-informed deep learning model to detect and correct motion artifacts in brain MRI, minimizing the risks of hallucination. Approach: We implemented a two-network framework — a motion predictor and a motion corrector — to identify k-space corruption and eliminate motion artifacts. Results: Our model outperformed existing methods, demonstrating superior artifact correction and soft-tissue contrast preservation, evidenced by higher PSNR and SSIM, and lower NMSE. Impact: Our physics-informed deep learning model markedly reduces motion artifacts in MRI scans, enhancing image quality. By minimizing the need for repeat scans, this method could significantly decrease healthcare costs and bolster the reliability of downstream MR imaging applications.