It may be possible to program computers to interpret pathology images to make cancer diagnoses more accurate. Harvard researchers are developing a machine-learning algorithm, using a “deep learning” approach, which interprets complex patterns observed in real-life data by building multi-layered artificial neural networks.1
“Artificial intelligence systems will be used to make pathological diagnoses of cancer more accurate and predictive, with the goal of helping oncologists to select the best treatment plan for each patient, based on the pathological characteristics of their disease,” said pathologist Andrew Beck, MD, PhD, director of bioinformatics at the Cancer Research Institute at Beth Israel Deaconess Medical Center (BIDMC), and associate professor at Harvard Medical School, in Boston, Massachusetts.
Dr Beck’s new approach was recently tested in a competition held at the annual 2016 International Symposium of Biomedical Imaging (ISBI). The test involved examining images of lymph nodes to decide whether they exhibited signs of breast cancer. The research team placed first in 2 separate categories, competing against private companies and academic research institutions from around the world.2
Continue Reading
In an evaluation in which researchers were given slides of lymph node cells and asked to determine whether they contained cancer, the team’s automated diagnostic method proved accurate approximately 92% of the time. This nearly matched the success rate of a human pathologist, who was 96% accurate.
“The testing data set consisted of 130 held out cases from 2 different institutions. On this test set, our system was 92% accurate by itself, and in combination with a pathologist was 99.5% accurate. We are actively testing on new cases to further determine its accuracy and to move it towards sufficient validation for clinical use,” Dr Beck told Cancer Therapy Advisor. “Next steps in advancing this technology include continuing to aggregate large datasets for training the AI-based systems, as well as developing systems methods of integrating these systems into the clinical workflow.”
The researchers started with hundreds of training slides, on which a pathologist had labeled regions of cancer and regions of normal cells. They extracted millions of these small training examples and used “deep learning” to build a computational model to classify them. The team then identified the specific training examples for which the computer is prone to make mistakes, and re-trained the computer using greater numbers more difficult training examples. The computer’s performance continued to improve.
Dr Beck said it was hoped that digitizing images and using machine learning could help pathologists work faster and make more accurate diagnoses for patients. Only recently, however, have improved scanning, storage, processing, and algorithms made it possible to pursue this automation effectively.
Sunati Sahoo, MD, of the Breast Pathology Services, and associate professor of pathology at the University of Texas Southwestern Medical Center in Dallas, said that this technology sounds promising, and may serve as an adjunct to current practices. But the software is not going to be cheap, nor easy to implement widely. “People at high volume centers may try this. It is going to be for all kinds of cancer, not just breast cancer. We don’t know the real value it adds, and it may add considerably to the workflow,” Dr Sahoo said in an interview Cancer Therapy Advisor.
RELATED: Watson Supercomputer Generates Personalized Cancer Treatments
Thaer Khoury, MD, professor of oncology in the Department of Pathology & Laboratory Medicine at Roswell Park Cancer Institute in Buffalo, New York, said that AI may play an important role in improving nomograms for predicting residual disease.
“There is emerging literature about the specific histological characteristics of tumor metastases in a lymph node to predict residual disease in the axilla, beyond using nomograms with encouraging results. This type of study, however, always suffers from inter- and intra-observer variability and reproducibility. Using image analyses would prevent this from happening, and would produce more consistent results,” Dr Khoury told Cancer Therapy Advisor.
References
- Wang D, Khosla A, Gargeya R, Irshad H, Beck A. Deep Learning Based Cancer Metastases Detection. https://www.dropbox.com/s/w9uxkxnoqifjnmw/Camelyon16_BIDMC_CSAIL.pdf?dl=0.
- ISBI challenge on cancer metastasis detection in lymph node. Camelyon 2016. http://camelyon16.grand-challenge.org/results.
