If you have ever had an MRI scan before, you will know how unsettling the experience can be. You are placed in a claustrophobia-inducing tube and asked to remain completely still for up to an hour, while invisible hardware, cripples and hugs around you look like a medical poltergeist. However, new research suggests that AI may help with this predicament by making MRI scans four times faster, getting patients in and out of the tube faster.
The work is a collaborative project called FastMRI between Facebook’s AI research team (FAIR) and radiologists at NYU Langone Health. Together, the scientists trained a machine learning model on pairs with low-resolution and high-resolution MRI scans, using this model to “predict” what latest MRI scans look like from just a quarter of the usual input data. This means that scans can be done faster, which means less hassle for patients and faster diagnosis.
“It’s a key step in incorporating AI into medical imaging,” said Nafissa Yakubova, a visiting biomedical AI researcher at FAIR who worked on the project. The edge.
The reason that artificial intelligence can be used to produce the same scans of lesser data is that the neural network has essentially learned an abstract idea of what a medical scan looks like by examining the training data. It then uses this to make a prediction about the final output. Think of it as an architect who has designed many banks over the years. They have an abstract idea of what a bank looks like, and so they can quickly create one last blueprint.
“The neural does not know the general structure of the medical image,” says Dan Sodickson, professor of radiology at NYU Langone Health, The edge. “In some ways, it is what we do fill in that is unique to this particular patient [scan] based on the data. ”
:format(webp):no_upscale()/cdn.vox-cdn.com/uploads/chorus_asset/file/21766684/AI_Blog_fastMRI_Desktop.jpg)
The fastMRI team has been working on this problem for years, but today they are publishing a clinical study in the American Journal of Roentgenology, what they say proves the reliability of their method. The study asked radiologists to make diagnoses based on both traditional MRI scans and AI-enhanced scans of patients’ knees. The study reports that when physicians are confronted with both traditional and AI scans, physicians make the same assessments.
“The key word here on which trust can be based is interchangeability,” Sodickson says. “We are not looking at anything quantitatively metric based on image quality. We say that radiologists make the same diagnoses. See fine the same problems. They miss nothing. ”
This concept is extremely important. Although machine learning models are often used to generate high-resolution data from low-resolution inputs, this process can often introduce errors. For example, AI can be used to scale low-resolution images from old video games, but people need to control the output to make sure it matches the input. And the idea of AI “imaging” a wrong MRI scan is obviously worrying.
However, the fastMRI team says that this is not a problem with their method. To begin with, the input data used to make the AI scans completely cover the target area of the body. The model for learning machines does not explain what a definitive scan looks like from just a few puzzle pieces. It has all the pieces it needs, just at a lower resolution. Second, the scientists created a control system for the neural network based on the physics of MRI scans. That means at regular intervals while performing a scan, the AI system checks that its output data matches what is physically possible for an MRI machine to produce.
“We just let the network create random images,” Sodickson says. “We require that every image generated by the process must have been physically realizable as an MRI image. We limit the search space in a way, and make sure everything is in line with MRI physics. ”
Yakubova says it was this particular insight that came only after lengthy discussions between the radiologists and the AI engineers, which enabled the success of the project. “Complementary expertise is key to creating solutions like this,” she says.
However, the next step is getting the technology into hospitals where it can actually help patients. The fastMRI team is confident that this can happen fairly quickly, perhaps in just a few years. The training data and model they have created are fully open access and can be incorporated into existing MRI scanners without new hardware. And Sodickson says researchers are already in talks with the companies that produce these scanners.
Karin Shmueli, who heads the MRI research team at University College London and was not involved in this study, said The edge this would be an important step to move forward.
“The bottleneck in taking something out of research to the clinic is often acceptance and implementation by manufacturers,” says Shmueli. She added that work like fastMRI was part of a broader trend that incorporates artificial intelligence into medical imaging which was extremely promising. “AI will definitely be more in use in the future,” she says.
