Magnetic Resonance Imaging (MRI) offers exceptional soft tissue contrast and high spatial resolution, all without the risks associated with ionizing radiation. MRI is capable of capturing a wide range of structural and functional information including blood flow, tissue perfusion, and myocardial strain making MRI a powerful tool for cardiac imaging. Unfortunately, achieving the spatial resolution necessary to extract this information often comes at the expense of temporal resolution. As a result, patient motion, such as cardiac or respiratory, can significantly degrade image quality, similar to the blurring that occurs when taking a photograph while the subject is moving. 

Cardiac MRI is highly susceptible to severe motion artifacts, which frequently necessitate repeat scans, may require patient sedation during imaging, and can result in images that are misleading or fail to provide the necessary diagnostic information. Automatically and accurately quantifying the degree of motion corruption in MRI scans would enable prompt identification of compromised images and support informed decisions about the necessity of rescanning, thereby preventing additional imaging sessions and reducing diagnostic errors.

In this work, we present a novel approach that simultaneously scores and corrects motion artifacts in cardiac MR images using a multi-task deep neural network. A key innovation of our method lies in the generation of training data: rather than relying on manually labeled images, we synthesize realistic cardiac and respiratory motion artifacts from high-quality, motion-free images by simulating their effects directly in k-space using a k-space swapping strategy. These synthetically corrupted images allow for controlled, large-scale training of a network that combines a U-Net for motion correction with a regression branch to estimate the severity of motion artifacts.Preliminary results indicate strong performance in both assessing and correcting motion-related artifacts in cardiac MRI.

The video of this presentation is available here.