“Looking back, there was clearly no tumor on the previous mammograms, but was there something in these very complex images that would hint at… a wrong development?” Barzilay asked. “It clearly didn’t just appear. Biological processes are in place to make a successful growth and it clearly impacts the tissue. So for a human who looks at it, it’s very hard to quantify the change, but a machine may look at millions of these images. This should really help them to look at these signs.”
Dr. Andy Beck, a pathologist at Harvard Medical School and Beth Israel Deaconess Medical Center, and Aditya Khosla, a computer scientist trained at MIT and Caltech, are tackling cancer diagnosis through images as well. They’re training computers to scour digital slides, and learn how to differentiate cells that are cancerous from those that aren’t.
They formed a startup, PathAI, last year after their technology won a competition in detecting breast cancer.
In the April 2016 challenge, an expert pathologist, charged with the same task as the computational teams, achieved an error rate of about 3.5 percent, Beck explained from PathAI’s headquarters in Cambridge, Massachusetts. Their team had an error rate closer to 7.5 percent, the best in the competition.
What was most interesting, Beck said, was putting the computer and pathologist together.
“The combination of human plus AI in this example reduced the expert’s error rate by 85 percent,” Beck said. “That was a really exciting finding.”
And because these trained computers get smarter with the more data they take in, the PathAI team’s technology improved over time. By November, Beck and Khosla’s system had surpassed the human expert.
Researchers are also using machine learning to make connections in data that people may not see. Joel Dudley’s team at Mt. Sinai in New York developed a system known as Deep Patient, scouring de-identified health data across the hospital system and combining information in multitudes of ways.
“One of the powerful aspects of deep learning is unsupervised feature learning,” Dudley explained, “meaning you don’t have to constrain upfront what you think is important for predicting something or modeling something.”
A physician or researcher focusing on type 2 diabetes, for example, may develop a model focusing on blood glucose or weight to try to predict who may be at risk for disease.
“But that then ignores all the other information in the health record that could be useful for predicting someone who’s at risk,” Dudley said. “So we use a deep learning approach where we could just pour in all the information we have on 5 million patients in our health system, from any test that’s ever been run on a patient.”
In results published last year in the journal Nature, Dudley’s team showed Deep Patient improved prediction of diseases from schizophrenia to cancer to severe diabetes.
MIT’s Barzilay was frustrated by medicine’s unsophisticated models as well. When she was going through treatment for breast cancer, she had questions about outcomes for different medicines for patients like her.
