Communications of the ACM
Scientists Train a Computer to Classify Breast Cancer Tumors
Researchers at the University of North Carolina Lineberger Comprehensive Cancer Center used a form of artificial intelligence to train a computer to analyze breast cancer images and then classify the tumors with high accuracy.
In "Image Analysis with Deep Learning to Predict Breast Cancer Grade, ER Status, Histologic Subtype, and Intrinsic Subtype," published in the journal NPJ Breast Cancer, researchers report they used machine learning, or deep learning, to train a computer to identify certain features of breast cancer tumors from images. The computer also identified the tumor type based on complex molecular and genomic features, which a pathologist can't yet identify from a picture alone. They believe this approach, while still in its early stages, could eventually lead to cost savings for the clinic and in breast cancer research.
"Your smartphone can interpret your speech, and find and identify faces in a photo," says the study's first author Heather D. Couture, a graduate research assistant in the UNC-Chapel Hill Department of Computer Science. "We're using similar technology where we capture abstract properties in images, but we're applying it to a totally different problem."
For the study, the researchers used a set of 571 images of breast cancer tumors from the Carolina Breast Cancer Study to train the computer to classify tumors for grade, estrogen receptor status, PAM50 intrinsic subtype, histologic subtype, and risk of recurrence score. To do this, they created software that learned how to predict labels from images using a training set, so that new images could be processed in the same way.
They then used a different set of 288 images to test the computer's ability to distinguish features of the tumor on its own, comparing the computer's responses to findings of a pathologist for each tumor's grade and subtype, and to separate tests for gene expression subtypes. They found the computer was able to distinguish low-intermediate versus high-grade tumors 82 percent of the time. When they had two pathologists review the tumor grade for the low-intermediate grade group, the pathologists agreed with each other about 89 percent of the time, which was slightly higher than the computer's accuracy.
In addition, the computer identified estrogen receptor status, distinguished between ductal and lobular tumors, and determined whether each case had a high or low risk of recurrence high levels of accuracy. It also identified one of the molecular subtypes of breast cancers—the basal-like subtype, which is based on how genes within the tumor were expressed—with 77 percent accuracy.
"Using artificial intelligence, or machine learning, we were able to do a number of things that pathologists can do at a similar accuracy, but we were also able to do a thing or two that pathologists are not able to do today," says UNC Lineberger's Charles M. Perou, the May Goldman Shaw Distinguished Professor of Molecular Oncology, professor of genetics, and of pathology and laboratory medicine in the UNC School of Medicine. "This has a long way to go in terms of validation, but I think the accuracy is only going to get better as we acquire more images to train the computer with."
The computer's ability to identify the basal-like subtype was exciting to researchers, and could have applications in cancer research. They also believe the technology could have applications in communities that do not have pathology resources as well as in helping to validate pathologists' findings.
"We were surprised that the computer was able to get a pretty high accuracy in estimating biomarker risk just from looking at the pictures," says UNC Lineberger's Melissa Troester, a professor in the UNC Gillings School of Global Public Health. "We spend thousands of dollars measuring those biomarkers using molecular tools, and this new method can take the image and get 80 percent accuracy or better at estimating the tumor phenotype or subtype. That was pretty amazing."
Deep learning technology has been used in a range of applications, including speech recognition and autonomous vehicles, Couture says.
"Humans can look at one or two examples of something, and be able to generalize when they see other objects," she says. "For example, chairs come in so many different forms, but we can recognize it as something we sit on. Computers have a much harder time generalizing from small amounts of data. But on other hand, if it you provide enough labeled data, they can learn concepts that are much more complex than humans can assess visually—such as identifying the basal-like subtype from an image alone."
The unique aspect of their work, researchers says, was that they were able to use the technology to see features of the tumors that humans cannot. They want to figure out what the computer is seeing, as well as to study whether the technology could predict outcomes.
"The computer extracted a lot of information from the images," Troester says. "We would like to test how well these features predict outcomes, and if we can use these features together with things like molecular data to do even better at giving patients a precise view of what their disease course looks like, and what treatments might be effective."
In addition to Couture, Perou, and Troester, other authors of the NPJ Breast Cancer study include Lindsay A. Williams, Joseph Geradts, Sarah J. Nyante, Ebonee N. Butler, J.S. Marron, and Marc Niethammer.
The study was supported by the University Cancer Research Fund through UNC Lineberger, the National Cancer Institute, the Breast SPORE Program, the Breast Cancer Research Foundation, and Susan G. Komen. The Tesla K40 used for this research was donated by Nvidia Corp.
Conflict of interest: Perou is an equity stock holder and board member for BioClassifier LLC, and has filed patents on the PAM50 subtyping assay.
No entries found