Two hedge fund quants have developed an algorithm that diagnoses heart disease from MRI images, beating nearly 1,000 other teams in one of the most ambitious competitions in artificial intelligence.
Tencia Lee, who recently left Los Angeles-based Crabel Capital Management to join a robotics start-up, said she and her partner Qi Liu, a former employee of Two Sigma hedge fund, had never worked before. with winning technology, called deep learning. .
Ms. Lee and Mr. Qi entered the competition in December and created a method that proved, in early tests, to be as effective as a cardiologist in analyzing images of the heart.
“People have been working on this for 15 years – I’m amazed at the kind of results that have come out of this competition in three months,” said Andrew Arai, chief of advanced cardiovascular imaging at the National Institutes of Health.
MRI scans are used to diagnose heart disease about 1 million times a year in the United States, Mr. Arai said, with cardiologists spending an average of 20 minutes on each image. This could make the algorithm a significant addition to a growing field of automated medical imaging, although it goes through rigorous formal testing before it can be adopted.
The winning entry used a convolutional neural network, a form of deep learning designed to mimic how vision works in animals.
Ms Lee said neither had worked with neural networks before and took software from GitHub, an online open source software repository, to solve the challenge. The main problem they faced was defining the problem with sufficient precision, she added. Then it was a matter of feeding the neural network with examples of cardiac MRI scans and letting it work out the solution.
The availability of such software meant that even complex problems could be solved by experts with more general training in data science, Ms. Lee said. “We’re both very used to working with large amounts of data and knowing where to look for problems,” she said. “It took all of my free time over a three month period. “
Recent advances in technology “allow us to do things with images that weren’t possible three or four years ago,” said Anthony Goldbloom, CEO of Kaggle, which hosted the National Science Data Bowl competition with the Booz Allen Hamilton consulting firm.
The changes include the development of specialized chips suitable for pattern recognition, called graphics processing units, or GPUs, and large increases in computing capacity, he added.
Open competitions have become an increasingly common way to solve difficult data science problems. Netflix hosted one of the biggest in 2009 with a million dollar competition to come up with an algorithm for making movie recommendations.
Kaggle, who runs about 50 contests a year on behalf of large corporations, said the cardiac imaging challenge was the toughest he’s ever faced. The chance to gain instant fame in their field, even more than the cash prize, explains the large number of people who have entered such contests, Goldbloom said.
Sander Dieleman, a doctoral candidate in neural networks who led a team at Ghent University that won the National Data Science Bowl last year, has since been hired by DeepMind, Google’s deep learning arm based in United Kingdom whose system recently beat the best human champion. to the board game Go.
Ms. Lee and Mr. Qi entered the competition individually and were ranked in the top 10 after a first round. They teamed up after Mr. Qi posted a message on Kaggle asking for help.
This story has been updated to clarify that Qi Liu no longer works for Two Sigma