AI Beats Pathologists in Predicting Survival in Brain Cancer

Roxanne Nelson, RN, BSN

March 21, 2018

New artificial intelligence (AI) software performed better than pathologists when it came to predicting overall survival in patients diagnosed with glioma, concludes a study published March 12 in the Proceedings of the National Academy of Sciences of the United States of America.

Predicting survival could be important for deciding on what treatment approach to pursue, the authors comment.

Led by Lee AD Cooper, PhD, a professor of biomedical informatics at Emory University School of Medicine and member of the Winship Cancer Institute in Atlanta, Georgia, the researchers developed a "computational approach" to predict the overall survival in patients with brain tumors, using microscopic images of tissue biopsy specimens and genomic biomarkers.

The software learned about survival from histologic images and created a unified framework that could integrate histology and genomic biomarkers and then predict time-to-event outcomes. The predictive ability of the software was ultimately more accurate than that of human pathologists.

"The eventual goal is to use this software to provide doctors with more accurate and consistent information. We want to identify patients where treatment can extend life," said Cooper in a statement. "What the pathologists do with a microscope is amazing. That an algorithm can learn a complex skill like this was an unexpected result."

He suggested that this study offers evidence of the profound impact that AI will have in medicine. "We may experience this sooner than expected," Cooper added.

Histology Important for 100 Years

The authors note in their paper that histology has been an important tool in cancer diagnosis and prognostication for over 100 years, and even though prognostication has been increasingly relying on genomic biomarkers, gene expression, and epigenetic modifications, histology continues to be an important tool in evaluating a patient's prognosis.

However, human assessments of histology are highly subjective and cannot be repeated, and thus, computational analysis of histology imaging has attracted a great deal of interest. Cooper and colleagues point out that advances in slide scanning microscopes and computing have allowed development of several imaging analysis algorithms for grading classification and identifying lymph node metastases in many types of malignancies.

Better Tools Needed

Deep convolutional neural networks (CNNs), they point out, have emerged as an important image analysis tool and "have shattered performance benchmarks in many challenging applications," the authors comment.

Using raw image data, the ability of CNNs to learn predictive features "is a paradigm shift that presents exciting opportunities in medical imaging," they add.

The authors developed a strategy called survival CNNs (SCNNs) that was able to provide highly accurate predictions of time-to-event outcomes using histology images.

Predicting Survival in Glioma

For this study, they evaluated the ability of this approach to predict overall survival in diffuse gliomas, a disease that has a wide range of outcomes. While prognosis and disease management depend on many factors, tumor grade typically will determine the treatment regimen. Grade III and IV tumors will generally be treated very aggressively, for example, while grade II tumors may even just be monitored in some cases.

Thus, there is a need to better discriminate more aggressive gliomas from those that are more indolent, they emphasize. These findings showed that SCNN worked as well as manual histologic grading or molecular subtyping for predicting overall survival. The authors also found that it was able to effectively determine within each molecular subtype and effectively performed digital histologic grading.

"Genomics have significantly improved how we diagnose and treat gliomas, but microscopic examination remains subjective," said Daniel J. Brat, MD, PhD, the lead neuropathologist on the study, who began developing the software while at Emory University and the Winship Cancer Institute, in a statement. "There are large opportunities for more systematic and clinically meaningful data extraction using computational approaches."

The authors conclude that while this tool has been used with gliomas, "validation of these approaches in other diseases is needed and could provide additional insights. In fact, our methods are not specific to histology imaging or cancer applications and could be adapted to other medical imaging modalities and biomedical applications."

The study was supported by US National Institutes of Health National Library of Medicine Career Development Award K22LM011576 and National Cancer Institute grant U24CA194362 and by the National Brain Tumor Society. Cooper leads a research project that is financially supported by Ventana Medical Systems; none of the other authors have disclosed any relevant financial relationships.

Proc Natl Acad Sci U S A. Published online March 12, 2018. Full text

For more Medscape Neurology news, join us on Facebook and Twitter


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.