Editor's Note: The video above is about ways to reduce the risk of breast cancer.
More than a million people every year in the U.S. get breast biopsies after an abnormality appears on a screening like a mammogram. But research has shown diagnoses from biopsies can vary between pathologists, and that variation may result in the over or under-treatment of patients.
Computer scientists at the University of Washington have developed artificial intelligence to improve the accuracy and consistency of those diagnoses.
When a person gets a biopsy, the tissue collected is put on a slide and examined by a pathologist. There are four potential outcomes when enlarged breast tissue is sampled.
A benign diagnosis means the tissue is just enlarged and there are no other issues. An “atypia” diagnosis means the cells are abnormal, but there are still no signs of cancer. A diagnosis of “ductile carcinoma in situ” or DCIS, means cancer cells are present, but they’re not behaving like cancer and breaking through cell walls. Finally, the most serious diagnosis – invasive cancer.
The good news is pathologists are good at detecting invasive cancer, however, the bigger challenge rests in determining cases that fall in the middle.
In the study, pathologists had 98% accuracy, compared to 91-94% from a computer using artificial intelligence when identifying invasive cancer versus non-invasive. But, when it came to distinguishing between atypia and DCIS, the computer had an accuracy of 88-89%, compared to 70% for pathologists.
Those in the middle diagnoses have important implications for future risks and patient care.
Roughly 10,000 patients a year get preventative mastectomies after a DCIS diagnosis, according to Dr. Joann Elmore, who is the director of the UCLA National Clinician Scholar Program and one of the researchers involved with the study.
“It can be very challenging, it’s easier to know when it’s cancer versus no cancer, but we also want to help women to know when they have these pre-invasive lesions and it’s these cases that are the hardest for the pathologist,” said Dr. Elmore. “It’s a really important differentiation because we provide different treatment for these different diagnoses.”
Dr. Constance Lehman is a professor of radiology at Massachusetts General Hospital. She has worked on computer-aided medical screening but was not involved with this research.
“This is a strong contribution to this area, but it’s a very early step,” said Dr. Lehman, noting she is cautiously optimistic. “It can be a challenging journey from early excitement and discovery to actually applying it in routine clinical practice and seeing the impact it has on patient outcomes.”
The artificial intelligence research at UW was led by Dr. Ezgi Mercan as part of her Ph.D. program. She used digitized images of 240 biopsies and then had three expert pathologists come to a consensus on the diagnosis to use as a baseline. From there, the images were shared with 87 other pathologists for diagnosis and compared to the results from the computer.
The study at UW was co-authored by Dr. Linda Shapiro, a professor of computer science. She explained most of the biopsies, about 68%, fell into the "atypia" or DCIS category. She said this could be part of why the computer did a better job of telling these apart.
“We only had 240 cases, they were carefully selected,” said Dr. Shapiro. “Machine learning does best when there are a lot of samples.”
She said future research incorporating more images could improve the performance of artificial intelligence.
“It’s an interesting era in medicine right now,” said Dr. Elmore, who is enthusiastic about the promise of machine learning. “There is profound potential for computers to help us. Computers are already interpreting EKGs, they’re already interpreting pap smears, they’re already assisting with computer-aided detection on screening mammograms. It’s an amazing time for computers and machine learning, there’s really great potential.”