|During MRI procedure, patients have to lie motionless inside the machine for a long time. / Photo by: Olena Yakobchuk via Shutterstock|
Facebook’s artificial intelligence researchers and New York University (NYU) School of Medicine collaborated in August on a project to make magnetic resonance imaging scan ten times faster than it takes to produce the result today.
Currently, it takes an hour or more to produce a quality MRI scan but it also means a long time for the patient to lie motionless inside the machine. This task may not seem difficult, but there are patients, especially children and those diagnosed with claustrophobia, who find the length of time a challenge. This is why the research team wanted to cut the time to just a few minutes.
Artificial intelligence for a faster MRI scan
Dubbed fastMRI, the team’s project involves producing a reliable MRI scan with the use of artificial intelligence and machine learning. First, they trained the AI with about 10,000 clinical cases as well as 3 million MRIs of the liver, brain, and knee. They ensured the public that no Facebook data was used in the project and the patient data was removed when they used the MRI scans. Their only purpose was to use the images to train the neural networks in recognizing the structure of MRIs and fill the missing views or parts. This is how the human brain works when the eyes view an incomplete image.
The team wanted the neural networks to bridge the gap but still maintain the accuracy of the MRI scans. In an update by science and technology magazine Popular Science, it shared that the tech giant and the New York University researchers are now testing the machine learning they developed to speed up the medical imaging procedure.
Testing machine learning
Radiologist Gina Ciavarra from NYU was among the study participants who tested the quality of AI scans against the traditional MRI. While she sat at NYU Langone Health to examine the MRI and X-ray scans, she determined that the grayscale images were from a patient’s knee. After studying the images, she detected a key problem and explained that the torn ligament is “definitely abnormal.”
She also made another evaluation of a knee scan, which she believes were created by an AI. Ciavarra and her team believe that by using AI in MRI scans, radiologists and computer scientists can speed up the medical exam. As a result, it reduces the need to anesthetize children who may have a hard time holding still while inside the scan machine. It also saves the hospital from spending more money. Facebook and NYU’s study is now to be submitted for academic review.
Benefits of using AI
The researchers believe that using artificial intelligence in such a way requires less data than the traditional approach but it can still create images that doctors can use to find out what is happening inside the body of the patient. A traditional MRI needs a certain amount of information. AI, on the other hand, needs less data, does fewer measurements, but gets the job done, just like traditional MRI scans. In short, it can still give the right picture.
Research scientist Larry Zitnic said that the machine learning-generated MRI needs to be accurate so that the surgeon and the radiologist can see the intel that they needed. He warned of the dangers if the scan misses a single tear in the ligament or if it invents something that is not actually there. Nevertheless, Facebook and NYU’s AI software was trained in such a way that it can correctly spin the data into MRIs.
|Using Artificial Intelligence in MRI scans needs less data, does fewer measurements, but gets the job done, just like traditional MRI scans. / Photo by: Laurent T via Shutterstock|
AI and radiology combined
They even conducted a blind test, where radiologists like Ciavarra reviewed the traditional MRI knee scans and images spun by AI to know if they can produce the same diagnostic information from the two scans. Zitnic described how radiologists had a hard time telling which among the images are from the AI and which one was done traditionally. Although there was added noise in the images, it still had no effect on the diagnostic value produced.
The use of artificial intelligence in medicine is indeed growing rapidly. In a 2018 study titled "Artificial intelligence in medicine: current trends and future possibilities," author Varun H. Buch from London-based home healthcare service provider Cera Care and Buch's team explained healthcare AI projects have even attracted more investments compared to other AI projects in other industries.
Further, chronic health conditions that are expected to benefit the most from AI and machine learning are diabetes (66 percent), heart disease (63 percent), cancer (63 percent), neurological diseases (56 percent), and infectious disease (46 percent). The use cases of AI and ML are the following: supporting clinical decision making (77 percent), extracting meaning from big data (66 percent), resolving operational inefficiencies (59 percent), enabling pop health management (52 percent), optimizing admin and clinical workflows (48 percent), increasing patient engagement (45 percent), advancing patient behavior change (43 percent), advancing personalized medicine (41 percent), and improving data integration from connected devices (41 percent). This is based on a survey by healthcare and technology platform Healthcare IT News.