Google’s DeepMind team showed that it can outperform trained radiologists in spotting cases of breast cancer, and that its artificial intelligence is capable of providing an independent, automated and immediate second opinion.
By using programs trained on 2D and 3D mammography images from nearly 30,000 women in the U.S. and the U.K., DeepMind’s system was able to identify those that had their cancer confirmed within the following year via a tissue biopsy or subsequent X-rays.
Among patients from the U.S., the AI cut the number of people incorrectly referred for further screening with a false-positive result by 5.7%—while also detecting 9.4% of potentially missed breast cancer cases. Published in Nature, the study said it surpassed the work of six independent physicians.
Additionally, Google’s researchers simulated its performance in automating the double-reading screening process employed in U.K. hospitals. By only bringing in a second human when the AI and the first clinician disagreed, they found the program was able to reduce the workload of the backup reader by 88%, while still maintaining the standard of care and providing on-the-spot feedback.
The study’s authors—some of whom had been recently reappropriated into Google’s wider healthcare efforts over the past year—said that the work of interpreting mammograms can still be challenging, with differences in patients’ breast density as well as variabilities among experts. The paper was the result of a collaboration with Cancer Research UK, Northwestern University and the NHS’ Royal Surrey County Hospital.
Clinical trials will still be needed to judge its performance in a real-world setting, as the system did not include all the different mammography technologies in use today, and most images were obtained from a single manufacturer’s system—according to an accompanying editorial in Nature by Etta Pisano, chief research officer of the American College of Radiology and a professor at Beth Israel Deaconess Medical Center at Harvard Medical School.
“In addition, it would be essential to develop a mechanism for monitoring the performance of the AI system as it learns from cases it encounters, as occurs in machine-learning algorithms,” Pisano wrote. “Such performance metrics would need to be available to those using these tools, in case performance deteriorates over time.”