The Australian Medical Association Limited and state AMA entities comply with the Privacy Act 1988. Please refer to the AMA Privacy Policy to understand our commitment to you and information on how we store and protect your data.

×

Search

×
27 Sep 2019

New research suggests artificial intelligence is able to interpret medical images and make diagnoses based on those images as well as human experts.

Through the use of deep learning algorithms, the potential for AI in healthcare has taken a leap forward, according to a review published in Lancet Digital Health.

AI advocates say the findings could help ease burdens on health systems around the globe.

But the review is based on only a small number of studies. And the experts themselves, have cautioned against suggesting that AI outperforms human capability.

Dr Xiaoxuan Liu, the lead author of the study from University Hospitals Birmingham NHS foundation trust, said the conclusion of the review had to be met with a reality check.

“There are a lot of headlines about AI outperforming humans, but our message is that it can at best be equivalent,” she said.

AI use in interpreting medical images, which relies on sophisticated machine learning, has shown promise in diagnosis of diseases, including cancers.

The researchers conducted the first comprehensive review of published studies on the issue and found AI to be on a par with humans in that area.

Because there is an abundance of poor quality study in the field, they focused on research papers published since 2012, which was a pivotal year for deep learning.

While the initial search offered more than 20,000 relevant studies, only 14 studies were reviewed because they were all based on human disease, reported good quality data, and tested the deep learning system with images from a separate dataset to the one used to train it. In those studies, the same images were also shown to human experts.

The researchers found that deep learning systems correctly detected a disease 87 per cent of the time and correctly gave the all-clear 93 per cent of the time, compared with 86 per cent and 91 per cent, respectively, for human experts.

 


Published: 27 Sep 2019