Office of Commercialization and Economic Development
Office of Technology Commercialization

Retrospective Artifact Correction of MRI through Unsupervised Deep Learning

Technology #20-0060

Questions about this technology? Ask a Technology Manager

Download Printable PDF

Researchers
Pew-Thian Yap
Managed By
Matthew Howe
Commercialization Manager 919.966.3929
Patent Protection
Provisional Patent Application Filed

Background
Structural magnetic resonance imaging (sMRI) is highly prone to motion artifacts, which is difficult to prevent in the case of pediatric patients. Improving image quality and removing artifacts retrospectively using supervised learning, requires the generation of paired data, consisting of images with and without artifacts of the same subject. Explicit generation of such images is impractical and challenging because it requires the generation of a multitude of potential artifacts. In consideration of these challenges, there is a need for improved MRI artifact correction.
 
Technology overview
Deep learning retrospect artifact correction (RAC) can improve image quality and evade the need for generating paired datasets by utilizing a cycle-consistency adversarial network that is trained with unpaired artifact-free and artifact-corrupted images. When using unsupervised training techniques and unpaired datasets, an adversarial network can be trained effectively to identify and eliminate artifacts with remarkable efficacy and can generate high-quality corrected images. Moreover, using an unsupervised training technique with an adversarial network will save significant time, effort, and resources since paired image sets are not needed for high-quality corrected images.
 
Researchers in the Department of Radiology and the Biomedical Research Imaging Center have developed an unsupervised deep learning retrospect artifact correction method that can generate high-quality corrected images. In this method, an adversarial network can be trained with an unsupervised training technique to transform artifact-corrupted MRI images to artifact-free MRI images. In this unsupervised training technique, the adversarial network can be trained to separate and encode MRI content and MRI artifact information from multiple unpaired sources. The unsupervised nature of this method allows for the removal of noise, streaking, and ghosting without explicit generation of artifacts from supervised training. Pediatric subjects were used to demonstrate the effectiveness of this method to remove artifacts and generate high-quality corrected images, without the need for specifying the artifacts.
 
Benefits

  • Does not require the explicit generation of artifacts for supervised learning
  • Paired MRI images or images from the same subject are not needed for retrospective artifact correction
  • Save time, effort, and resources compared to supervised methods
  • Efficient removal of artifacts
Applications
Retrospective artifact correction via unsupervised deep learning can be used to improve image quality in MRI for pediatric patients by utilizing a disentangled cycle-consistency adversarial network.