Newly developed AI capable of identifying prostate cancer with 'near-perfect accuracy'
PITTSBURGH, July 28, 2020 (StudyFinds.org) — Human error can be charming in an endearing kind of way, but no one appreciates mistakes when it comes to a topic as serious as cancer. On that note, researchers from the University of Pittsburgh developed a new artificial intelligence program with the most accurate record to date when it comes to recognizing prostate cancer.
“Humans are good at recognizing anomalies, but they have their own biases or past experience,” says senior author Dr. Rajiv Dhir, chief pathologist and vice chair of pathology at UPMC Shadyside and professor of biomedical informatics at UPitt, in a release. “Machines are detached from the whole story. There’s definitely an element of standardizing care.”
What separates this AI from the rest of the robotic pack? Dr. Dhir and his team “fed” their program images from over a million parts of tissue slides extracted from prostate cancer patient biopsies. Then, the AI program was tested on 1,600 different slide images collected from 100 suspected prostate cancer patients.
The AI performed incredibly well on that test. Results show 98% sensitivity and 97% specificity at finding and identifying prostate cancer. Those stats are much higher than scores recorded by previous cancer-detecting algorithms.
Moreover, this is the first algorithm that does far more than just detect cancer. This program also scores well in categories including tumor grading and sizing, and assessment of surrounding nerve invasion by cancer cells.
The AI even detects cancer in six slides that had slipped by a human pathologist.
While all of this is very promising, the study’s authors caution that the AI program isn’t quite ready to fully replace human doctors just yet. For example, regarding those six slides that went unnoticed by a human doctor, that pathologist could have just seen enough evidence to warrant a cancer diagnosis before coming upon those particular slides.
Still, Dr. Dhir says the algorithm, at the very least, can serve as a great failsafe.
“Algorithms like this are especially useful in lesions that are atypical,” Dr. Dhir comments. “A nonspecialized person may not be able to make the correct assessment. That’s a major advantage of this kind of system.”
Of course, this project only focuses on prostate cancer. While an entirely new algorithm would have to be trained for each type of cancer, the research team is optimistic their results can be recreated with other cancer variations.
The study is published in The Lancet Digital Health.