Artificial intelligence helps detect ovarian cancer early and accurately
Ovarian cancer is difficult to diagnose, particularly in its early stages, when survival rates are much higher. Because there is no consistently reliable screening test to detect ovarian cancer, most women are diagnosed with the disease when it’s in an advanced stage.
Source: Greg Slabodkin
However, researchers at Brigham and Women’s Hospital and Dana-Farber Cancer Institute have developed a non-invasive diagnostic test using artificial intelligence for the accurate detection of true cases of early-stage disease. Results of their study were published online this week in the journal eLife.
Brigham and Women’s Hospital
By combining next generation sequencing with artificial intelligence, researchers have created a novel blood test based on serum microRNAs—small, non-coding pieces of genetic material that help control where and when genes are activated—for the early diagnosis of ovarian cancer.
Also See: Georgia Tech makes machine learning algorithm available to fight cancer
According to Kevin Elias, MD, the lead author and assistant professor in BWH’s Department of Obstetrics and Gynecology, the problem with current early detection tests such as ultrasound and CA-125 protein testing is that they have high false positive rates for ovarian cancer and do not have a meaningful impact on survival rates.
“In the study, a large part of getting it to work was evaluating a range of different artificial intelligence programs—we went through actually 33 different types of models,” says Elias, who notes that the team sequenced the microRNAs in blood samples from 135 women to create a “training set” with which to train a computer program to look for microRNA differences between cases of ovarian cancer and cases of benign tumors, non-invasive tumors and healthy tissue.
Researchers leveraged large amounts of microRNA data in this machine learning approach to develop different predictive models and although many of the models performed well, the one that most accurately distinguished ovarian cancer from benign tissue was a neural network model, reflecting the complex interactions between microRNAs.
“When we train a computer to find the best microRNA model, it’s a bit like identifying constellations in the night sky. At first, there are just lots of bright dots, but once you find a pattern, wherever you are in the world, you can pick it out,” he adds.
Having confirmed the accuracy of the model, researchers applied the model to 859 patient samples to measure the sensitivity and specificity of the model. What they found was that 100 percent of abnormal results using the microRNA test actually represented ovarian cancer, while using ultrasound fewer than 5 percent of abnormal test results would be ovarian cancer.
“The model significantly outperformed CA-125 and functioned well regardless of patient age, histology or stage,” states the article, noting that the primary advantage of the neural network over CA-125 was avoiding false positives. “Among 454 patients with various diagnoses, the miRNA neural network had 100 percent specificity for ovarian cancer.”
In addition, researchers deployed the final model using the microRNA diagnostic test to predict the diagnoses of 51 patients presenting for surgical care in Poland. In this population, 91.3 percent of the abnormal test results were ovarian cancer cases—a very low false positive rate—and negative test results reliably predicted absence of cancer about 80 percent of the time, which is comparable to the accuracy of a Pap smear test.
“The key is that this test is very unlikely to misdiagnose ovarian cancer and give a positive signal when there is no malignant tumor. This is the hallmark of an effective diagnostic test,” said Dipanjan Chowdhury, PhD, Chief of the Division of Radiation and Genomic Stability in the Department of Radiation Oncology at Dana-Farber Cancer Institute.
While the test was applied to ovarian cancer, Elias contends that it is easily adaptable to other kinds of cancer. “There’s no reason that we can’t take the same steps and apply it to other types of malignancies,” concludes Elias. “If you’re going to analyze microRNAs as a biomarker, this is the technique that you want to use.”