Artificial intelligence can be more effective than experienced doctors at identifying breast cancer abnormalities in ultrasound images, research by New York University Abu Dhabi and a US medical centre has shown.
The study follows other findings in which AI has been shown to outperform clinicians in medical diagnosis and it could signal a major shift in how the technology is used in hospitals.
One of the scientists behind the latest study, Dr Farah Shamout, an assistant professor of computer engineering at NYU Abu Dhabi, said AI was likely to have “quite enormous” an effect in medicine.
“AI is showing high potential in improving the performance of certain tasks like diagnosis and predicting prognosis,” she said.
“Once it replaces the traditional techniques that we’ve been using that have been shown to underperform compared to AI, it will have a great effect in terms of improving patient outcomes.”
Published in the journal Nature Communications, the new study looked at how effective a neural network - a computer system that carries out tasks in a way similar to how the brain does - was at identifying malignant lesions in ultrasound images. These are images produced by scanning a tissue or organs with sound at a frequency higher than can be heard by people.
The neural network was trained using more than 280,000 ultrasound images from more than 140,000 patients examined at NYU Langone, a New York-based medical centre, between 2012 and 2019.
It was found to perform better on average than 10 certified radiologists at highlighting potentially cancerous lesions in ultrasound images.
The study also found, however, that a combination of AI and analysis by radiologists was the most effective way to check ultrasound images, being better than AI alone or doctors on their own.
Another key finding was that AI could help to cut the number of false positives, which are cases where breast cancer was suspected, when it was actually not present. False positives can lead to unnecessary biopsies, which waste money and can be unpleasant for patients.
“The use of the network can improve the specificity of cancer diagnosis and reduce false positive findings, which can lead to unnecessary biopsies,” said Dr Shamout.
“We’re not only improving the diagnosis of cancer, but also improving the confirmation there’s no cancer, which is also important for patients.”
The new study, co-written by NYU Langone researchers, is one of a number indicating that AI is as effective or better than clinicians at diagnosing disease.
Previously, researchers in Chicago found that AI could detect lung cancer in chest scans earlier than doctors did. Similarly, a study in London found the technology was better than doctors at detecting breast cancer from mammograms, which are images of the breast produced using low-energy X-rays.
Although the new research indicates that AI may be more accurate than radiologists, it is not ready to replace them.
The study involved analysing ultrasound images and patient data from New York alone, so the method needs to be tested elsewhere.
The researchers are looking to collaborate with hospitals in other countries, including the UAE, to enable them to test the neural network’s performance with data from additional patient groups. They would also like to test the AI with other diagnostic methods, such as mammograms.
“We’re definitely interested in transferring it to the clinic, but this would involve clinical trials; I think this needs a bit more time,” said Dr Shamout.
In a project being run in collaboration with Cleveland Clinic Abu Dhabi, Dr Shamout is also using AI to predict urinary tract infections, which could allow the use of antibiotics to be more targeted at patients who really need them.
