Machine Learning Could Speed Up Radiation Therapy For Cancer Patients

"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 Richard Radke, assistant professor of electrical, computer, and systems engineering at Rensselaer. Radke is leading a team of engineers and medical physicists to develop a "machine learning" algorithm that could cut hours from the process.

A subfield of artificial intelligence, machine learning is based on the development of algorithms that allow computers to learn relationships in large datasets from examples. Radke and his coworkers have tested their algorithm on 10 prostate cancer patients at Memorial Sloan-Kettering. They found that for 70 percent of the cases, the algorithm automatically determined an appropriate radiation therapy plan in about 10 minutes.

"The main goal of radiation therapy is to irradiate a tumor with a very high dose, while avoiding all of the healthy organs," Radke said. He described early versions of radiation therapy as a "fire hose" approach, applying a uniform stream of particles to overwhelm cancer cells with radiation.

IMRT adds nuance and flexibility to radiation therapy, increasing the likelihood of treating a tumor without endangering surrounding healthy tissue. Each IMRT beam is composed of thousands of tiny "beamlets" that can be individually modulated to deliver the right level of radiation precisely where it is needed.

But the semi-automatic process of developing a treatment plan can be extremely time-consuming -- up to about four hours for prostate cancer and up to an entire day for more complicated cancers in the head and neck, according to Radke.

A radiation planner must perform a CT scan, analyze the image to determine the exact locations of the tumor and healthy tissues, and define the radiation levels that each area should receive. Then the planner must give weight to various constraints set by a doctor, such as allowing no more than a certain level of radiation to hit a nearby organ, while assuring that the tumor receives enough to kill the cancerous cells.

This is currently achieved by manually determining the settings of up to 20 different parameters, or "knobs," deriving the corresponding radiation plan, and then repeating the process if the plan does not meet the clinical constraints. "Our goal is to automate this knob-turning process, saving the planner's time by removing decisions that don't require their expert intuition," said Radke.

The researchers first performed a sensitivity analysis, which showed that many of the parameters could be eliminated completely because they had little effect on the outcome of the treatment. They then showed that an automatic search over the smaller set of sensitive parameters could theoretically lead to clinically acceptable plans.

The procedure was put to the test by developing radiation plans for 10 patients with prostate cancer. In all 10 cases the process took between five and 10 minutes, Radke said. Four cases would have been immediately acceptable in the clinic; three needed only minor "tweaking" by an expert to achieve an acceptable radiation plan; and three would have demanded more attention from a radiation planner.

Radke and his coworkers plan to develop a more robust prototype that can be installed on hospital computers and evaluated in a clinical setting. He hopes to see a clinical prototype in place at Memorial Sloan-Kettering in the next few years. The researchers also plan to test the approach on tumors that are more difficult to treat with radiation therapy, such as head and neck cancers.

In a related project, Radke is collaborating with colleagues at Boston's Massachusetts General Hospital to create computer vision algorithms that offer accurate estimates of the locations of tumors. This automatic modeling and segmentation process could help radiation planning at an earlier stage by automatically outlining organs of interest in each image of a CT scan, which is another time-consuming manual step. Learn more about this project here: http://news.rpi.edu/update.do?artcenterkey=134.

The research is supported by the National Cancer Institute and the Center for Subsurface Sensing and Imaging Systems (CenSSIS) at Rensselaer, which is funded by the National Science Foundation. Renzhi Lu, a graduate student in electrical engineering at Rensselaer, also contributed to the research.