Overdiagnosis is a growing problem in the medical community, especially in the field of breast pathology. Diagnosing patients with breast cancer when they don’t actually have it can lead to stress and unnecessary treatments.
To address this problem, a team of UW and UCLA researchers used artificial intelligence to create a diagnostic support system for medical experts.
This research, funded by the National Institutes of Health, was the dissertation of Ezgi Mercan, a doctoral student in the Paul G. Allen School of Computer Science & Engineering, who is currently working as a researcher at Seattle Children’s Hospital. She worked closely with Linda Shapiro, professor of the Allen School, and Dr. Joann Elmore, professor of the David Geffen School of Medicine at UCLA.
“I joined the team as an image analysis researcher and I was lucky to be in the right place at the right time to have the resources to work on a topic I was personally really passionate about,” Mercan said.
Breast pathology has a wide spectrum, ranging from benign tissue to atypia, to ductal carcinoma in tissue (DCIS), and invasive cancer. A DCIS diagnosis shows there are non-invasive cancerous cells in the breast tissue, which means they do not grow into normal tissues, whereas atypia does not indicate cancer.
Mercan said the team oversampled atypia and DCIS cases in their diagnostic study because past studies have shown that there is a high rate of disagreement among pathologists when differentiating between the two.
In the study, the team developed a novel image descriptor and used machine learning to model various diagnoses. This new image feature was trained to recognize diagnoses ranging from benign to invasive cancer based on a dataset of 240 breast biopsy whole slides.
The accuracy of this automated system was compared to that of the ground truth diagnoses given by three expert pathologists, and the diagnoses made by 87 other pathologists across the country. Mercan said humans are able to diagnose invasive cancer cases and differentiate benign cases, but the performance of their system was superior with atypia and DCIS cases.
Shapiro said the system they developed is also distinct from other automated systems.
“This particular system differs from others coming out currently in that the diagnostic results come from carefully designed features, rather than from deep learning,” Shapiro said.
One of the most challenging parts about this project was the many years it took to gather the data, which ultimately revealed how much physicians cared about maintaining a high quality of care in their field.
“We appreciated the busy pathologists who agreed to help out and participate in the study, as they spent up to 20 hours involved in this study without any financial compensation,” Elmore said.
Although the research was unique in and of itself, Mercan said the opportunity to work with two other women on her dissertation committee was special for her, as gender disparity is a huge issue in the field of computer science.
“It was a rare opportunity to have role models and mentors on my committee that supported my personal and professional development as well as my research,” Mercan said. “By simply observing their dedication and work ethic, I came to expect more from myself and eventually achieved more.”
Since the publication of their work last August, Shapiro’s students have been furthering this research by adapting the system to diagnose skin cancer with the help of dermatopathologists. This is much more challenging because these images differ from those of breast cancer.
“We are using some of the basic methods from the breast cancer work, but they lead to a very different kind of structure,” Shapiro said. “We are in the middle of the work so there is no diagnosis yet, but we’re hopeful.”
In the age of technology we are living in today, Elmore said it’s important to recognize that the use of artificial intelligence isn’t new to medicine.
“We already use AI to interpret pap smears, to interpret EKG’s, and we already have computer-aided detection programs in mammography, thus, AI is already in place in the practice of medicine,” Elmore said.
As the team works to make this system publicly available, Mercan said that it is a step toward developing an actual clinical system.
“A number of recent articles published by interdisciplinary research teams like ours prove that both medical and tech communities realize the potential of AI in diagnostic medicine,” Mercan said. “We can expect to see AI slowly integrated into clinical systems.”
Reach reporter Shannon Hong at firstname.lastname@example.org. Twitter: @shannonjhhong
Like what you’re reading? Support high-quality student journalism by donating here.