FAIR, a Facebook AI research team, is a project that aims to reduce the burden on patients by realizing MRI scan speeds 10 times faster by using AI with NYU Langone Health. ) In progress. The research team has announced a technology that can improve the shortcomings of fastMRI and improve the quality of existing MRI scan images.
Research results on FastMRI are already public, and code and datasets are also being released on GitHub to build a reference model for the broader community to engage in FastMRI. On February 25 (local time), the research team announced a technology that solves the fast MRI problem through deep learning and improves the image quality of the entire MRI scan.
According to the research team, one of the tasks to perform an accurate MRI scan on the original data using deep learning is the banding and streaking artifacts in the scanned image. This artifact, which can be called noise among MRI scan images, is visible to trained experts, but not easy for beginners. The research team explained that while evaluating images of high-speed MRI scans, they discovered that artifacts that went horizontally could significantly degrade the image quality and obscure the disease.
To solve this problem, the research team created a deep learning learning model that retrieves low data from MRI scans and generates accurate MRI images without artifacts through hostile learning. Hostile learning is often compared to the relationship between banknote counterfeiting and police. In other words, the counterfeiter creates a counterfeit bill that is as close to the real thing as possible, and the police determine whether it is real or fake, but when the police’s ability to judge it increases, the counterfeit technique also improves to trick it. It is a similar structure.
In the case of FastMRI, the research team set the hostile learning goal to predict artifact pattern propensity. Since the hostile model and the MRI image reconstruction model are trained at the same time, the reconstruction model is provided until the artifact disappears, while the accuracy of the hostile model’s detection of artifacts has been continuously improved.
In the image, the left is an MRI scan image accelerated by AI by a normal MRI scan, and the right is an image created by hostile learning. On the right, you can see that the extra oblique line disappears, so it is a clearer image.
As MRI accelerated by AI is becoming a big issue with artifacts, the technology released this time has the potential to be the first step in enabling fast MRI to be used in clinical settings. In addition, as advanced MRI scanners have reported a tendency to easily create artifacts, it is revealed that this technology can help improve. Related information can be found here .