NUH’s Spine AI quickens discovery of spine issues

The Singaporean hospital expects the volume of scans to surge amidst an ageing population.

Singapore’s National University Hospital (NUH) has developed an artificial intelligence (AI) tool that cuts the time it takes to interpret MRI (magnetic resonance imaging) scans of the lower back, a needed boost as the city-state faces the challenges of an ageing population.

The hospital’s deep learning tool Spine AI, which helps detect lumbar spinal stenosis, reduces the interpretation time to three minutes from 10, said James Hallinan, senior consultant at NUH’s Department of Diagnostic Imaging at NUH.

This allows radiologists to quickly review flagged areas and produce accurate reports.

Lumbar spinal stenosis, a narrowing of the spinal canal in the lower back that leads to nerve compression, is common among the elderly. It can cause significant discomfort, limiting mobility and quality of life, according to NUH.

The hospital expects the volume of scans to surge, with one in four Singaporeans hitting more than 65 years by 2030, compared with one in 10 in 2010.

"The model detects the area of interest in the image and classifies the degree of narrowing, where the nerves could be squeezed, as normal, mild, moderate, or severe,” Hallinan, a radiologist, said. “We're looking at saving probably 466 hours a year.”

He said junior radiologists have found it helpful because they can see what their superior is reviewing, including his edits. “This allows for discussion and the use of spine AI bounding boxes to clarify my observations.”

To ensure the model can detect spinal abnormalities, NUH doctors worked with data scientists from the National University of Singapore’s School of Computing, using about 450 spines and 18,000 images.

“We used labels from experienced radiologists to invent the most accurate model,” Hallinan said.

Given Singapore’s racially diverse population, Hallinan said they also obtained data sets from Saudi Arabia for external validation. The team also worked with Siemens Healthineers so the algorithm could be deployed with a user-friendly interface.

“The other great thing about working with Siemens is we could roll this out internationally,” he said. “They have bases across the world, including Europe, which means we could potentially conduct prospective studies internationally.”

The initial feedback for the product, which is still in its trial phase, has been positive, with radiologists appreciating Spine AI being only one click away from use, Hallinan said.

The system’s accuracy has been touted, though a bigger study is needed to validate these findings, with plans to expand the technology to other institutions in Singapore, he added.

Beyond its application in spinal imaging, NUH plans to expand AI’s role in other areas. The hospital is looking into chest X-ray AI algorithms to identify cases that require immediate attention or those suitable for discharge.

“Then there’s rapid AI for detecting stroke,” Hallinan said. “We're also looking at mammogram models, several in-house ones, and then commercial ones as well.”

Hallinan said AI does not replace and only supports radiologists. “With algorithms, we need to make sure that they're safe, and that there's an overview from radiologists. We must strive for efficiency and potential cost savings whilst prioritising safety."