Home - Emerging Tech - Neural Networks May Enable Faster MRI Scanning
Research Brief

Neural Networks May Enable Faster MRI Scanning

Neural networks trained to fill in information gaps may be able to create high-quality MRI images from less data.

By
Jill Sakai, Contributor
Friday, September 14, 2018

SHARE: 

The project, called fastMRI, is a collaboration between the Facebook Artificial Intelligence Research group and the New York University School of Medicine. The goal, as described in a press release, is to construct diagnosis-quality images using faster scans and less data.

MRI scans provide better detail of soft tissues than other types of medical imaging, such as X-ray or CT scanning, but are relatively slow due to the large amount of data they collect. FastMRI seeks to accelerate the scanning process by capturing less raw data. Then, similar to how the human visual system constructs a complete view of the world from incomplete information, artificial neural networks will be trained to fill in the remaining information gaps to produce detailed images. One challenge the project will address is the need for extremely high accuracy in the recreated images to allow accurate and thorough diagnoses.

The researchers are using a data set compiled by the NYU School of Medicine that includes about 3 million anonymized MRI images of the knee, brain and liver from 10,000 clinical cases. They will compare traditional and artificial intelligence-based reconstructions of the scans to train the neural networks and assess performance.

Traditional MRI scans can take anywhere from 20 minutes to more than an hour, depending on the body region. Reducing required scan times has the potential to make MRI scanning accessible to more people, including those who are unable to lie still for the prolonged time required for a typical scan. A shorter scan time could also help relieve scheduling backlogs in places with limited MRI access. The project partners suggest that rapid MRI might even replace X-ray or CT scanning in some applications, reducing patient radiation exposure.

Both the image data set and the AI models and evaluation metrics used will be made publicly available to encourage broad adoption and additional research developments, according to the project partners.