British radiologists were no match for a new machine learning method that was able to identify fractures with almost 20% more accuracy and confidence.

A research team at the University of Bath created two convolutional neural networks to identify and classify hip fractures from X-ray scans. When compared to human radiologists, the CNNs did the job 19% better.

Classifying a fracture is crucial for selecting the right type of surgery to perform, and is also timely, as such procedures should ideally take place within 48 hours to avoid adverse outcomes and mortality risks. But no standard process exists to determine who is right to classify these issues, an orthopedic surgeon or radiologist specializing in musculoskeletal disorders.

The team hopes that their technology can standardize this process and speed up the diagnostics process by alleviating the bottleneck of 300,000 radiographs that remain unreported in the U.K. for over 30 days. In addition, they are seeing it as a helpful solution for the shortage of radiologists in the U.K. and elsewhere, and for meeting rising demands for radiology services, according to The Washington Post.

"The volume of medical imaging performed routinely is growing and exceeds the capacity of human observers to report on findings. There is tremendous potential for AI systems to highlight the presence of pathology to human observers. It should be remembered, though, that if the AI systems have not been trained to look for a particular pathology, they will not be able to detect it," professor Richie Gill, lead author of the paper and co-director of the Center for Therapeutic Innovation at the University of Bath, told HCB News.

According to the American College of Radiology, an estimated 30% of radiologists use AI in their work and even more are considering adopting it. This could be especially helpful for hip fractures, say the researchers. In 2019, 67,671 hip fractures were reported to the U.K. National Hip Fracture Database. In the U.S., an estimated 300,000 hip fractures happen every year, with the number expected to rise to more than 500,000 by 2040, reported The Washington Post.

At the same time, the number of radiographs performed increased 25% in the U.K. between 1996 and 2014. This rising demand often makes it impossible for radiologists to interpret and deliver results in a timely manner. The NHS also has a shortage of nearly 2000 consultant radiologists on hand. And the U.S. is facing these same problems as well.

The team trained and tested the solution with 3,659 hip X-rays that were classified by two experts. It showed overall accuracy of 92% compared to the 77.5% shown by the clinicians. While accuracy varied depending on the fracture, the researchers said the algorithms showed “an impressive, and potentially significant” ability to classify them.

"We are looking at both expanding the capabilities of this system for classifying hip fractures into more refined subtypes and using this type of approach for other imaging modalities such as CT and PET. Three-dimensional imaging contains a huge amount of data, and conventionally, it is reported by looking at 2D views of the data set; this is an area in which AI methods may be able to exploit all the available information," said Gill.

He adds, however, that despite outperforming humans in this study, the AI technology is not meant to, and unlikely to, replace radiologists. "The machine learning systems at the moment can only identify pathologies they have been trained on. Perhaps more importantly, a human observer can instantly tell if the correct image has been presented. For example, if a knee radiograph is wrongly labelled as a hip radiograph in the imaging system and then passed on for reporting for hip pathology, the human will immediately realize it is the wrong image. This is not so straightforward for the AI system, and considerable effort would be needed for AI systems to be able to establish that the correct image is being used for a given reporting stream."

The study was funded by the Arthroplasty for Arthritis Charity. The NVIDIA Corporation provided the Titan X GPU that carried out the machine learning.

The findings were published in Nature Scientific Reports.