Software developed for mine dectection may held doctors spot cancer cells.
When Duke engineer Larry Carin took on a project to develop software for the U.S. Office of Naval Research, the goal was to improve the detection of undersea mines. Now, that same software may end up helping doctors screen patients for cancer, too.
The surprising link stems from a shared challenge in identifying mines and cancer cells: Both rely on computer systems to sift through piles of data and turn up signs of danger. And in both cases, the machines may lack enough information to make the correct decision.
To improve the performance of the Navy's robotic mine-hunting systems, Carin designed active-learning software that helps the computers classify unknown objects. Using information theory, the software asks a human to provide labels for objects the system can't recognize.
It turns out the software also works remarkably well for classifying images of human cells. Medical researchers at the University of Pennsylvania applied the algorithms developed by Carin to an automated image-analysis program commonly used by doctors to identify cancer cells and significantly improved the program's performance. The enhanced toolkit lets physicians label fewer cell samples because the algorithm automatically selects the best set of examples to teach the software.
"The results are spectacular," says Carin, William H. Younger Professor and chairman of electrical and computer engineering. "This is not a typical Navy transition, but it is a transition to a very important medical tool used at hospitals around the world. There is a real chance this may save lives in the future. This could be a game-changer for medical research."