Algorithm promises faster radiotherapy planning

February 13, 2007

Software that learns how to automatically fine-tune plans for conducting radiotherapy promises to eliminate the need for manual adjustments, speeding up the process needed to administer radiation to cancer patients, according to its developers at the Rensselaer Polytechnic Institute. The algorithm is being groomed to automatically determine acceptable radiation plans in minutes.

Software that learns how to automatically fine-tune plans for conducting radiotherapy promises to eliminate the need for manual adjustments, speeding up the process needed to administer radiation to cancer patients, according to its developers at the Rensselaer Polytechnic Institute. The algorithm is being groomed to automatically determine acceptable radiation plans in minutes.

The need for such software is palpable, according to Richard Radke, Ph.D., an assistant professor of electrical, computer, and systems engineering at Rensselaer.

"Intensity-modulated radiation therapy (IMRT) has exploded in popularity, but the technique can require hours of manual tuning to determine an effective radiation treatment for a given patient," said Radke, whose group described its solution in the Feb. 7 issue of Physics in Medicine and Biology.

In preliminary tests conducted on 10 prostate cancer cases, the algorithm automatically determined in 10 minutes or less an appropriate radiation therapy plan for seven.

Machine learning, a subfield of artificial intelligence, uses algorithms that analyze examples, then search databases for similar relationships. It holds particular promise for the management of sophisticated technologies such as IMRT, whose radiation beams are composed of thousands of tiny "beamlets" that must be individually adjusted to deliver precise doses of radiation.

Interactively constructing the treatment plans that orchestrate this modulation of doses, however, can be extremely time-consuming. Technologists can spend up to four hours to develop a plan for treating prostate cancer and an entire day developing plans to treat cancers in the head and neck, according to Radke.

The exact locations of the tumor and healthy tissues are identified using CT images. Radiation doses are assigned to specific target areas, while the need to avoid collateral radiation damage to healthy tissues must also be weighed. This is now done by manually determining the settings for up to 20 different parameters.

"Our goal is to automate this knob-turning process, saving the planner's time by removing decisions that don't require their expert intuition," Radke said.

In its first outing, the algorithm constructed radiotherapy plans based on data describing 10 patients with prostate cancer. Each radiation plan required between five and 10 minutes. Of the 10 plans developed, four would have been immediately acceptable and three needed only minor tweaking. The remaining three would have demanded more attention from a radiation planner, Radke said.

The Rensselaer team is now working on an enhanced version that can be evaluated clinically. In the meantime, the team will expand the approach to plan treatment for head and neck cancers. The group is also collaborating with Massachusetts General Hospital to create computer vision algorithms that can automatically outline organs of interest in CT images, further speeding the radiation planning process.