Human doctors once again fell short of artificial intelligence in a test to accurately diagnose breast cancer, adding yet more evidence that AI-aided diagnostics may soon be commonplace. Researchers at the University of Washington and UCLA created a system that distinguished between a pair of conditions that human doctors often struggle to identify correctly. The results are reminiscent of an effort by Google to detect metastatic breast cancer using AI, which the company says is 99 percent accurate. Rival tech giant Microsoft has several projects that use algorithms to fight cancer and work with Seattle-based Adaptive Biotechnologies on a system that uses AI to diagnose multiple diseases from a single blood test. The company claims that another project aimed at spotting early-stage lung cancer can also outperform doctors.
While the researchers’ system is far from autonomous, it may become a tool to help doctors be more accurate, said Linda Shapiro, a computer science professor at the UW. “We’re billing it as something like computer-aided diagnostics, where it can make suggestions and show pathologists what it’s thinking and why,” she said. Shapiro said the breakthrough was thanks to the efforts of Ezgi Mercan, a doctoral student at the University of Washington who invented a new way to describe cancerous structures so that they could be understood by machine learning algorithms. Mercan based the approach on conversations with Dr. Donald Weaver, a pathologist who showed Mercan his methods for detecting breast cancer.
The results were published today in the journal JAMA Network Open. The algorithm outperformed human pathologists when tasked with identifying ductal carcinoma in situ from atypia — two conditions that are two commonly misdiagnose conditions. Accuracy rates were 89 percent for the system and 70 percent for the pathologists. The computer-based approach had similar accuracy to the group of 87 doctors in differentiating cancer from non-cancer tissues. Distinguishing breast atypia from ductal carcinoma in situ is important clinically.
Very challenging for pathologists. Sometimes doctors do not even agree with their previous diagnosis when they are shown the same case a year later,” Dr. Joann Elmore, a professor of medicine at the David Geffen School of Medicine at UCLA, said in a statement. One hurdle standing between U.S. researchers and a reliable AI system for cancer diagnostics is the lack of available data, given strict rules governing patient records. In China, researchers built a highly accurate method to diagnose childhood diseases that were made more relaxed, in part, by the country’s more relaxed data-sharing laws. To create the dataset needed to train their algorithms, the UW and UCLA researchers relied on Dr. Jamen Bartlett, who labeled dozens of complex images over nine months. Sachin Mehta, a doctoral student at UW, also contributed to the project.