Turning the signal-
to-noise ratio on its ear
Scientists Get Patent for "Noisy" Algorithms That Could Improve Mammography
January 06, 2010
by Brendon Nafziger
, DOTmed News Associate Editor
Scientists at Syracuse University recently received a patent for a series of algorithms that they say can cut down the number of false positives in digital mammography by more than a third. And they do it, paradoxically, by adding "noise."
In general, scientists hate "noise," messy interference that obscures the signal, or whatever target they're looking for. But new research shows that some noise actually makes signals clearer, and the finding could have a big impact on detecting breast cancer.
The patented algorithms developed by the researchers depend on something called stochastic resonance. This phenomenon was first used to account for why small changes in the earth's orbit every 100,000 years can trigger massive changes in climate -- the ice ages.
Essentially, scientists reasoned that some signals too faint to reach a threshold for detection could actually pole-vault over if boosted by "noise." This is called random, or stochastic, interference. In the case of the ice ages, the "signal" is orbital variation of the earth, which by itself is too weak to cross the "threshold" of worldwide freezing. But coupled with "noise," the fluctuating bombardment of solar energy randomly heating the planet, and the signal has the capacity to cross the threshold and set off an avalanche of events that can lead to global winters, or so the theory goes.
"Basically, in our daily experience, if a signal is more noisy, it will be degraded," says Hao Chen, Ph.D., a research assistant professor in the electrical engineering and computer science department of the university, and one of the holders of the patents. "A noisy image looks much worse," he says.
"But for some other systems, under certain conditions, and there are lot of restrictions; but for some systems, if you make the input signal more noisy, then the output performance would be better. That's quite counterintuitive. Usually, we do a lot of de-noising before you process the data. But in this case, we do the opposite, we add noise to signal and make it more noisy, and somehow it improves the system performance," he says.
That is the case with certain nonlinear systems, like the earth's climate, but it also appears to be true with mammography.
Eliminating false positives
Thanks to several grants from the Air Force Office of Scientific Research, Chen and his colleagues Pramod Varshney and James Michels, also professors at Syracuse, did the research that led to the 2006 filing of the patent. But it wasn't until 2009 that, with doctoral student Renbin Peng, they published in the Journal of Selected Topics in Signal Processing how their system can help find breast cancer.
What Peng and the team did was to look at digital mammograms of breasts with tiny calcifications measuring some 0.3 millimeters in size. They then ran a series of publicly available algorithms known to detect and mark breast cancer on digital mammograms. When they matched these markings against ones made by their "noise" adding algorithms, they found the stochastic resonance-based program was much less likely to mistake benign growths for deadly tumors.
With their detectors, there was a 36 percent reduction in false positives, even though the sensitivity -- the ability to detect real cancers -- was the same or better than for other systems.
"It's just a prototype," says Chen, "but we demonstrated some potential. We believe it can be applied to a broad spectrum of applications, such as cochlear implants or systems to improve touch."
As for offers from OEMs, Chen says he's open, and some inquiries have already been made through an office at the university.
"We were hoping that someday we can collaborate with some companies to actually make a product that will be beneficial for society," he adds.