Portland, Ore. - Researcher Harold Szu, working at the Naval Research Laboratory (Arlington, Va.), has turned advanced target recognition technology into what he hopes will be a method for containing the spread of SARS. His 200-channel infrared body scanner, equipped with an unsupervised-learning Lagrange constraint neural network (LCNN) algorithm, can see into the body at any depth and resolve local hotspots-a sign of diseased cells.
Szu plans to test the system on real patients at Thermal-Scan Inc. (Detroit). After that he wants to take the equipment to Toronto, where the SARS epidemic is still in progress.
"If they will let us scan real SARS patients, then my unsupervised neural network could learn the distinctive patterns of heat generated inside in local areas [in those affected with SARS], so that we could probably detect SARS [in other individuals] before it becomes contagious," he said.
The technique uses infrared cameras designed for satellite imagery but focused for a six- to eight-foot scan. Infrared scans have been in use in the medical community to detect breast cancer, since rapidly reproducing cells also give off more heat than normal cells. But the technique has been very limited because only one spectral band is used. The military scanning equipment that Szu is using operates on multiple wavelengths and also has two imaging systems that can pinpoint the distance of a hotspot.From space to cells
Szu and James Buss of the Naval Research Lab designed the LCNN algorithm to locate relevant information in the high-volume real-time data produced by the system's 200 channels. The original application was for remote sensing of the earth from satellites, but the researchers found that the same equipment and algorithm could be used to image the human body at infrared wavelengths.
"We can see the infections inside the body long before the elevated blood temperature works its way out to your thermometer," Szu said. The unsupervised-learning neural network reads the deep infrared images starting at the surface of the skin, proceeding all the way through flesh and bone. The neural-network analysis program then compares the images with historical data on file. Gradual changes in temperature can indicate growing cancer, while rapid changes in heat images indicate an infection caused by a virus or bacteria.
A trial run of the equipment last year showed a striking contrast to current infrared detection techniques. Szu's teams' imaging system had a much higher resolution and was far more successful at locating cancer cells than single-wavelength scans.
A fundamental problem with heat imaging in the body is the random variation in heat from one point to another: It is very difficult to filter out the random signal to detect a distinctively hotter spot. Resolution is also low.
Szu's imaging system gets around the problem by using two cameras operating at different wavelengths along with the neural-network algorithm. That lets the system isolate the important variations in a signal that would indicate a malignant cell or infection.
Szu believes that the new technique will turn out to be far more effective at detecting breast cancer than current mammography systems and that it would be an ideal tool for scanning individuals for the presence of infectious disease.