Skip to main content
Premium Trial:

Request an Annual Quote

Emory, Northwestern Developing Glioma Predictor That Could Surpass Pathologists' Analyses


NEW YORK (360Dx) – Researchers at Emory and Northwestern universities have developed artificial intelligence software that analyzes data from tissue biopsies to predict the survival of patients diagnosed with glioma.

According to the researchers, their approach may be more accurate and objective than the predictions of clinicians with years of specialized training.

Describing the approach in the Proceedings of the National Academy of Sciences, the researchers said earlier this month that it presents "an innovative approach for objective, accurate, and integrated prediction of patient outcomes."

Gliomas are often fatal within a couple of years of diagnosis, but with accurate diagnosis and appropriate care, some patients can survive more than 10 years, Lee A.D. Cooper, the study's lead author and a professor of biomedical informatics at Emory University School of Medicine, said in an interview.

"Our idea was to have an algorithm that could look at histology, or some combination of histology and genomics, and predict the expected outcome for a patient," Cooper said. "We've shown that by combining deep-learning technology with survival models used by biostatisticians, you can build a machine that extracts important visual patterns from slide images and do a very good job at predicting the expected survival."

Doctors currently use a combination of genomic tests and microscopic examination of tissues to predict how glioma's will behave clinically or respond to therapy, Cooper said. Genomic tests are extremely accurate, and pathologists "are quite good" at predicting rates of survival, but grading tissue biopsies for survivability after a glioma diagnosis brings subjectivity and variability in results, he said.

"It's very difficult to extract reliable measurements from tissue images because they are so complex," Cooper said. "However, people are showing that deep-learning methods that are adaptive can do a good job of making predictions from these images."

"It's critically important to improve diagnosis based on tissue biopsies and genetic testing," Melissa Bondy, a professor of medicine, epidemiology, and population sciences at Baylor College of Medicine, said in an interview. "As things move forward with genetics, this group is taking analysis of survivability to another level using artificial intelligence."

The group is developing a different technology that begins to better classify tumors, she said. "That's really the beginning of where we will be able to go with big data and machine language, and being able to look at cancers in a different way in combination with genetics and histologic subtypes."

Bondy, who is also associate director of cancer prevention and population sciences in the Dan L Duncan Comprehensive Cancer Center at Baylor, led an international consortium of researchers from 14 cancer centers in 2017 that conducted a large study of malignant brain tumors looking for genetic markers of glioma.

Although the genetic biomarkers they identified could be developed to evaluate glioma risk, the new artificial intelligence research looks at glioma diagnostics through the lens of survival.

"It takes the clinical aspects of diagnosis to a different level," Bondy said. She noted that the technique is not yet "ready for primetime, where pathology is going to immediately incorporate it, but hopefully in time it will."

The Emory and Northwestern research could eventually lead to a commercial diagnostic test, but it could also be "a number of years" before the group is able to validate and prepare a product for market introduction, Cooper said.

As a next step, the group intends to conduct a large prospective study that evaluates patient outcomes. The group has "the infrastructure and expertise needed" to conduct the study, and it is discussing options with potential funding partners, he said.

The researchers "would probably try to partner with a diagnostic company" that has the infrastructure and knowledge to introduce a commercial product to the market, Cooper said. "There's a lot of activity in the digital pathology space, and quite a number of startups and established companies that are doing similar things" to his group, he noted.

In their study, the researchers used deep-learning techniques on microscopic images of brain tumor tissue samples to train the software to learn visual patterns associated with patient survival. The study used public data produced by the National Cancer Institute's Cancer Genome Atlas project to develop and evaluate the algorithm.

Genomic testing provides reliable results, but pathologists still need to review tissue to determine whether the cancer is aggressive and how aggressive it is, Cooper said.

Different pathologists sometimes provide different interpretations for the same case, the researchers said. The way tissue is handled prior to testing can produce variability in results, and that's likely to continue, Cooper said. Further, the variability that comes from differences in interpreting what's on a slide can impact important clinical decisions, such as whether a patient enrolls in an experimental clinical trial or receives radiation therapy as part of their treatment, Copper added.

Daniel Brat, the lead neuropathologist on the study said in a statement, that genomics has "significantly improved how we diagnose and treat gliomas, but…there are large opportunities for more systematic and clinically meaningful data extraction using computational approaches."

Currently chair of pathology at Northwestern University Feinberg School of Medicine, Brat began developing the software with Cooper while he worked at Emory University and the Winship Cancer Institute. In 2015, to evaluate the genetic content of gliomas in the brain, Brat with colleagues published the results of a study in the New England Journal of Medicine that leveraged brain tumor information from The Cancer Genome Atlas, which brings together genomic and clinical data with pathology images.

A growing body of evidence prompted the World Health Organization in 2016 to change its classification of tumors of the central nervous system that included gliomas defined by not just histology but also molecular features.

"That reduced a lot of subjectivity, but grading still involves histology" and subjectivity, Cooper said. After a genomic classification has been completed, "a pathologist looks at the slide to determine whether the disease is aggressive and how advanced it seems to be. That's what we primarily wanted to target with the new approach — reducing the subjectivity in grading."

The eventual goal is to use this software to provide doctors with more accurate and consistent information that could identify patients who could extend their lives through treatment, Cooper said. "This is more evidence that AI will have a profound impact in medicine, and we may experience this sooner than expected."

Cooper noted that despite the promise and accuracy of genomic testing, the study shows that "there is still a lot of value in patterns in histology tissues, and that we can apply algorithms that makes histology and digital pathology an important part of cancer research and clinical care in the future." 

Ichiro Nakano, a professor of neurosurgery and academic neurosurgeon at the University of Alabama at Birmingham, said he sees other areas that need improvement. Advances in molecular profiling in the past decade "has given us a better idea of the behavior of the tumor" after a biopsy, he said.

"We definitely need better and more accurate diagnosis methods prior to doing a tissue biopsy," he said, because surgeons try to avoid doing biopsies unless they are completely necessary. "Improving imaging technology is one way to achieve this."