Above are two CT scans of the body taken when the breath is held, and then expelled. While doctors can accurately aim radiation beams, they often irradiate a larger-than-necessary area since breathing, eating habits, or movement can change the location of the tumor and make it difficult to pinpoint.

Patients undergoing radiation treatment eventually may have a smaller amount of healthy tissue exposed during the procedure.

Rich Radke, assistant professor of electrical, computer, and systems engineering, is collaborating with colleagues at Boston's Massachusetts General Hospital to create computer vision algorithms that offer more accurate estimates regarding the locations of tumors in patients undergoing treatment.

Radke says that during radiation therapy, determining the precise position of a tumor on a given day can be difficult. "For example, the prostate moves around depending on what you had to eat that day, or if your bladder is full or empty. In the case of a lung tumor, your breathing pattern affects where the tumor is at any given time." The tumor itself also may be changing size and position between treatments. To compensate, therapists are often forced to irradiate a wider region of the body than is strictly necessary.

To overcome these challenges, Radke is working to "teach" computers to recognize the shapes and positions of tumors in diagnostic images - such as CAT scans - and creating templates for radiotherapy treatment. He envisions a real-time computer system that reads an image, analyzes it with computer vision algorithms, displays the estimated tumor position to a radiation oncologist, and sends the accurate location directly to the radiation beam.

Currently, it takes about 20 minutes for a trained radiation therapist to outline the position of a tumor prior to each daily treatment. The new method will reduce that time. "Our algorithm presents its best guess to a trained person who checks it over and will nudge the contours if there's a mistake." He adds, "Our algorithm will learn from these corrections so that it's less likely to make the same mistake next time. The more the algorithm is used, the better it should get."

Radke says he has good preliminary results with 2-D images of the prostate. "The next step is to take our algorithm to 3-D, and improve the fitting algorithm so that it makes fewer mistakes." In the following year, he hopes to be able to track images of tumors that move according to breathing patterns.

While his work currently focuses on prostate tumors, Radke begins a new one-year exploratory project this fall to develop algorithms that will apply to other parts of the body.

The research is supported by Rensselaer's Center for Subsurface Sensing and Imaging Systems (CenSSIS). Radke is working with colleagues Badri Roysam, professor of electrical, computer, and systems engineering, and Daniel Freedman, assistant professor of computer science. •


Originally published in Rensselaer Magazine, Fall 2002

Published September 1, 2002