|Machine vision is also starting to enter into everyday life. Many digital cameras now have a feature that will automatically locate and highlight people’s faces. Some even have a feature that can automatically tell if everyone’s eyes are open and if they’re smiling, and the camera will actually delay taking the picture for a fraction of a second until everyone in the shot is open-eyed and grinning.
“It’s cool to see algorithms from computer vision conferences making their way, slowly but surely, into the real world,” Radke says. “It’s definitely getting transitioned out of labs and into new applications in ways that make a difference to the public.”
As cameras become standard features in most cell phones, he envisions that machine vision technology in today’s high-end digital cameras will trickle down into future phones. Radke also sees machine vision playing a larger part in our automobiles and driving habits.
“I can easily imagine an active windshield that uses some sort of subtle visual cue to highlight pedestrians, or point out that another vehicle is on a collision course with your car,” he says. “Or maybe your car is automatically figuring out where the sun is, proportionate to your eyes, and it puts up a black spot so you don’t have to put on sunglasses or flip down your visor.”
Another major area of Radke’s research program involves medical image analysis, especially in the context of radiation therapy. Every time a patient undergoes a computerized axial tomography (CAT) scan, it results in a series of 3-D images. In preparation for administering radiation therapy, doctors scroll through these images, locate the afflicted organs, such as the prostate, and then circle the tumor in each slice. Radke has worked to automate this process by designing machine vision techniques for the same task, using organ models built from many expert outlines. The net time saved for each scan may only be a few seconds, but when multiplied by the tens of thousands of doctors who handle hundreds of these images every week, it becomes substantial.
Radke’s groundbreaking work in biomedical image processing is supported by CenSSIS, the Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems, a research center funded by the National Science Foundation and dedicated to developing new technology for identifying objects that are somehow obscuredfrom cancerous tumors under the skin to pollution plumes underground.
Radke was particularly interested in the computer vision and machine learning problems related to intensity modulated radiotherapy, or IMRT, a new technology for cancer treatment. Working over the last several years with Rensselaer undergraduate and graduate students, and collaborating with medical physicists at Memorial Sloan-Kettering Cancer Center in New York City and Massachusetts General Hospital in Boston, Radke’s team has investigated IMRT-related problems for both breast and prostate cancer.
Radke and his students also have designed algorithms that can play a critical role in helping a medical physicist develop a radiation treatment plan once the tumor and critical organs have been outlined, a complex task that involves calculating the different angles and intensities at which radiation is beamed into the body. The goal, again, is trimming time.
“It can take hours for an expert to develop a radiation treatment plan, but if we can use our algorithms to provide a viable or near-viable plan in just a few minutes, that saves hours of a planner’s time, allowing him or her to do other important stuff,” he says. “We’re not trying to replace anyone with a computer, or shift the responsibility of patient care from doctors to a machine. We’re just trying to save people time by speeding up processes that are routine enough to be automated.”
Radke finds this work especially rewarding, and he continues to look for more applications in cancer treatment for the techniques he’s developed. “It feels good to see the impact, to talk with the people in the trenches who are actually treating patients and hear that the work I was doing has some real bearing,” he says.