AI-Powered MRI Quality Control and Artifact Correction for Multi-Site Studies

AI-Powered MRI Quality Control and Artifact Correction for Multi-Site Studies Abstract Magnetic resonance imaging (MRI), while commonly employed to investigate brain structure and function, is susceptible to imaging artifacts caused for instance by eye and head motion, hemodynamic changes, and mag- netic field inhomogeneities. Identifying images with questionable quality, correcting for artifacts, and discarding unusable images reduce biases in subsequent analyses, and therefore help avoid erroneous conclusions. In this project, we will develop computational tools, powered by artificial intelligence (AI), for image quality assess- ment (IQA) and retrospective artifact correction (RAC), catering to large-scale studies involving multiple imaging centers. In Aim 1, we will develop a deep learning framework for automatic, objective, fast, and accurate IQA of structural MRI data. IQA is typically performed via visual inspection by MRI technologists and can be time-consuming, subjective, and error-prone. Our method will complete IQA in milliseconds with high sensitivity and specificity. It will be designed to be easily trainable with a small amount of annotated data via semi-supervised learning. It can be used immediately after each scan to determine whether the acquired data are usable and whether re-scan is necessary. It can also be used in post-acquisition data curation to automatically identify usable images from huge image datasets. We will also develop a deep learning framework to rapidly generate derivatives, such as cortical surfaces, for inspection to ensure sufficient image quality for downstream analyses. In Aim 2, we will develop a novel RAC method to remove artifacts without requiring modification of sequences or mounting of motion markers. Existing RAC neural networks are typically trained in a supervised manner, requiring the scanning of the same subjects motionless and with deliberate motions. In contrast, our RAC method will be based on unsupervised learning, simply requiring the user to provide for training a set of clean and corrupted images, which can be from different individuals. Our method is therefore practical with considerably better adaptability. In Aim 3, we will develop techniques that will allow the methods developed in Aims 1 and 2 to be adapted to and optimized for different imaging centers, MRI scanners, and imaging protocols without requiring explicit exchange of imaging data across sites, which often requires data sharing agreements between institutions. The techniques developed in this aim will help overcome data sharing challenges for site-optimized IQA and RAC. Successful completion of this project will allow image quality assurance to be done automatically in large-scale, multi-site, and longitudinal studies, and increase the amount of usable data for improving statistical power.