Welcome to the Qinsun Instruments Co., LTD! Set to the home page | Collect this site
The service hotline


Related Articles

Product Photo

Contact Us

Qinsun Instruments Co., LTD!
Address:NO.258 Banting Road., Jiuting Town, Songjiang District, Shanghai

Your location: Home > Related Articles > Facebook and New York University use AI to accelerate MRI scanning speed by four times

Facebook and New York University use AI to accelerate MRI scanning speed by four times

Author:QINSUN Released in:2024-02 Click:29

According to the foreign media The Verge, a new study recently showed that using artificial intelligence can accelerate the scanning speed of magnetic resonance imaging (MRI) by four times. This work is a collaborative project between Facebook's AI Research Team (FAIR) and radiologists at New York University Langley Medical Center, called fastMRI.

Scientists trained a machine learning model together on a pair of low resolution and high resolution nuclear magnetic resonance scans, using this model to "predict" the final appearance of the nuclear magnetic resonance scan from only a quarter of the usual input data. This means that scanning can be completed faster, which means patients have less trouble and diagnosis is faster.

"This is an important stepping stone for integrating artificial intelligence into medical imaging," Nafissa Yakubova, a visiting researcher at FAIR Biomedical Artificial Intelligence who participated in the project, told The Verge.

Artificial intelligence can be used to generate the same scan from a small amount of data because neural networks have essentially learned an abstract idea of understanding what medical scans look like by examining training data. Then, it utilizes this to predict the final output.

"Neural networks know the overall structure of medical images," Dan Sodickson, a professor of radiology at the Langley Medical Center at New York University, told The Verge. "In some aspects, what we are doing is filling in the unique features of this specific patient (scan) based on data."

The fastMRI team has been studying this issue for many years, but on Tuesday they published a clinical study in the American Journal of Roentology, which they said proves the credibility of their method. This study requires radiologists to diagnose patients based on traditional MRI scans and artificial intelligence enhanced knee scans. The research report states that when faced with traditional and AI scans, doctors make identical evaluations.

"The key word that can be based on trust here is interchangeability," Sodickson said. "We are not looking at some quantitative indicators based on image quality. We are saying that radiologists make the same diagnosis. They discover the same problem. They won't miss anything."

This concept is extremely important. Although machine learning models are often used to create high-resolution data from low resolution inputs, this process often introduces errors. For example, artificial intelligence can be used to enhance the low resolution images of old video games, but humans must check the output to ensure it matches the input. The idea of AI imagining the wrong MRI scan is clearly worrying.

However, the fastMRI team stated that this is not a problem with their methods. Firstly, the input data used to create AI scans completely covers the target area of the body. Machine learning models do not rely solely on a few puzzle pieces to guess what the final scan result will look like. It has all the fragments it needs, but the resolution is low. Secondly, scientists have created an inspection system for neural networks based on the physical principles of nuclear magnetic resonance scanning. That is to say, during the process of creating a scan, the artificial intelligence system will check periodically whether its output data matches the data that may be physically generated by the MRI machine.

"We're not just letting the network create any arbitrary image," Sodickson said. "We require that any image generated through this process must be a physically achievable MRI image. We have to some extent limited the search space to ensure that everything is consistent with MRI physics."

Yakubova said that it was this particular insight that was realized after a long period of discussion between radiologists and artificial intelligence engineers, which led to the success of this project. "Complementary expertise is key to creating such solutions," she said.

However, the next step for scientists is to bring this technology into hospitals and truly help patients. The fastMRI team is confident that this can be achieved quite quickly, perhaps in just a few years. The training data and models they create are completely open and can be incorporated into existing MRI scanners without the need for new hardware. And Sodickson said that researchers are already in negotiations with the companies that produce these scanners.

Karin Shmueli, the head of the MRI research team at University College London (who was not involved in the study), told The Verge that this will be a crucial step in advancing the research.

"The bottleneck of bringing something from research into clinical practice is often the adoption and implementation of manufacturers," Shmueli said. She added that work like fastMRI is part of a broader trend to incorporate artificial intelligence into medical imaging, which is very promising. "Artificial intelligence will definitely have more applications in the future," she said.