AUSTIN, Texas — Advances in artificial intelligence are leading to machine learning algorithms that can analyze and interpret human biopsy samples and theoretically detect cancer, all in the blink of an eye. That is, unless it makes incorrect predictions when it comes to close calls — a consideration expressed in a new perspective piece published in The New England Journal of Medicine.
“One of my concerns is that physicians will blindly accept findings made by these algorithms as truth, as being better than findings by human pathologists, without taking into account weaknesses in the technology,” said co-author Adewole “Ade” Adamson, M.D., MPP, assistant professor of internal medicine at Dell Medical School at The University of Texas at Austin and a dermatologist at UT Health Austin. “I fear some clinicians think machine learning is magical, and it’s just not,” said Adamson, who is also a member of Dell Med’s Livestrong Cancer Institutes.
The Fundamental Flaw of Being Human
Central to the reliability of machine learning in cancer diagnosis is the human foundation on which the algorithms are based.
“In the case of many cancers, the gold standard is diagnosis under a microscope by a pathologist. Unfortunately, this is a flawed approach. There’s a wide variability between and among pathologists with regard to early cancer diagnosis. Because machine learning algorithms train on data annotated by a flawed gold standard, using these algorithms may not get us any closer to the truth of what is truly cancer,” Adamson said.
“How on Earth are you going to teach an algorithm to make definitive conclusions when pathologists often can be uncertain or even get it wrong?” Adamson asked.
Particularly with early stage cancer, Adamson said the level of threat is difficult.
The Issue of Overdiagnosis
Overdiagnosis is another potential problem resulting from AI, Adamson says. Currently when making a cancer diagnosis, pathologists usually examine a relative handful of biopsy slices of tissue under a microscope, taken from different sections of the sample.
Whereas a human pathologist’s diagnosis might decline in accuracy as the person tires, machine learning algorithms can consistently examine infinitely more samples for possible signs of cancer. As a result, the algorithms may increase the likelihood of finding minor abnormalities labeled as “cancer” that never would have harmed the patient, resulting in overdiagnosis.
The Case for Uncertainty
Adamson and his co-author — H. Gilbert Welch, M.D., MPH, with Brigham and Women’s Hospital in Boston — call for a deeper investigation of machine learning in cancer diagnosis before relying heavily on its abilities.
They also argue machine learning tools that analyze tissue samples should offer a third, intermediate choice of uncertainty, not just “cancer” or “not cancer.”
This intermediate choice could require additional analysis by a pathologist to make final determinations about the presence of cancer. Otherwise, overreliance on machine learning could lead to misdiagnoses — or overdiagnoses that lead to unnecessary treatment, Adamson said.
The researchers applaud what algorithm analysis does bring to exam rooms: faster and more consistent inspections of biopsy samples. Adamson adds that though machine learning algorithms promise to deliver quicker and more consistent diagnoses and improve patient care, these algorithms may not be fully reliable in getting closer to the truth.
o: 512-495-5062, c: 347-419-2657