In the summer of 2015, Santhosh Subramanian was 17 and about to enter his senior year at Northview High School. He was spending the break not hanging with friends but living in Bethesda, Maryland, where he was doing an internship at the National Cancer Institute’s biostatistics branch.
One day he learned something alarming: CT scans were notoriously unreliable in distinguishing malignant tumors from benign ones. In fact, the rate of false positives—tumors that turned out to be harmless—was a whopping 96.4 percent. So tens of thousands of people every year were going through the stress of additional tests, such as biopsies, and waiting periods while their doctors watched the abnormality to see if it grew.
Around the same time Subramanian learned about this, his high school buddy Sanjit Kumar was learning that his father had a tumor growing on his forehead. The two teens had been fast friends since their AP calculus class their junior year. They lived in neighboring subdivisions, and both had parents working in the software industry. Plus, they shared a tendency to dream up ideas for companies.
The elder Kumar went through three scans before doctors determined the tumor was benign. But the experience left his son taking stock of his life in a way that most of us don’t do until our 40s, if ever. “I didn’t want my dad to end up passing, and if he did, for him to think of what I could have done,” Sanjit Kumar says. “I wanted him to be able to think of what I did do.”
The two friends started out with a simple question: Is there a better way to diagnose cancer using artificial intelligence and predictive analytics? Put another way, would it be possible to design software that could examine and learn from past diagnoses to predict future outcomes? Their answer, one that someday might make these two teenagers millions, is DiaScan, a risk assessment tool that takes CT scans and quantifies tumor characteristics to look for patterns that differentiate between benign and malignant growths.
“Images are an array of numbers,” Subramanian says. “We process the image such that [the software] isolates the tumor and extracts [its] features.” Those features are compared to those of other tumors in its memory banks to see how the numbers match up. The more comparisons, the better it will get at determining the odds of malignancy. “The beauty of machine learning and data science is that as we add more data, we’re able to find beautiful patterns,” Subramanian says.
While the technology is still under wraps until its planned release later this year, Subramanian and Kumar liken it to the facial recognition tools used by Instagram and Facebook. “The stuff we’re doing isn’t complicated,” Kumar says. “But bringing the technology to healthcare is what makes us special.” He points out a paradox of the medical industry. “When you think about it, there’s a huge disconnect between medicine and technology.” Especially where it comes to actually putting big data to use. Social media companies are on top of it. But hospitals? Not so much.
An early version of their software earned them a ticket to a local pitch-off competition held by startup and tech news site TechCrunch, where they won third place and the audience choice award. They also met Jim Schwoebel, cofounder of local startup accelerator CyberLaunch, which invested $20,000 of pre-seed funding in the technology. At CyberLaunch they also met Sam Franklin, vice president of data science at marketing agency 360i, who serves as CyberLaunch’s mentor in data science. “I felt in my gut that these were guys I needed to work with because the problem they were taking on was so impactful,” he says. “I lost my father very early in my life to brain cancer. Most of my family members who have passed away have passed away due to cancer, so it’s something that tugs at me personally.” While the software right now is designed to help detect lung cancer, he sees its long-term potential as much bigger. “This could be a generalized tool in our battle against cancer both here in the States and—ideally—globally.”
“In the end,” Kumar says, “it’s just applying machine learning and computer vision to the medical industry. It doesn’t necessarily have to be cancer. It could be any disease a radiologist would [use images to diagnose].”
Currently the pair are using the CyberLaunch funds to develop a pilot version of their software while they scout out $250,000 in seed funding for additional product development. They’ve also begun talks with CT scanner manufacturers, medical software companies, data organizations, and hospitals for licensing agreements, pilot studies, and partnerships.
In the meantime, they’re off to separate colleges—Subramanian to Berkeley, where he’s majoring in computer science, and Kumar to Georgia Tech, where he’s majoring in computer engineering. Life in the on-campus dorm sometimes brings into focus just how different Kumar’s priority list is from those of his peers.
“Sometimes I’ll see my friends and think, ‘Man, I just want to be 18,’” he says. “Other times I wish I was in my 20s so I could be taken more seriously.”