Crowdsourced AI learns to target lung tumors for radiation

(Reuters Health) – In many parts of the world there are not enough radiation oncologists to design and deliver radiation treatments for lung cancer patients, but that gap could one day be filled with the help of artificial intelligence, researchers suggest in a new study.

In a novel approach to the problem, the authors turned to crowdsourcing to help them develop a computer algorithm that would take over some of the duties of an experienced radiation oncologist.

“Lung cancer is a major global health problem,” said the study’s lead author, Dr. Raymond Mak, an associate professor of radiation oncology at Harvard Medical School, the Dana Farber Institute and Brigham and Women’s Hospital in Boston. “It’s the number one cancer killer in the world. Close to a million patients will need radiation therapy at some point.”

Currently, radiation oncologists determine where to send the beams of radiation by drawing the outlines of the tumor and surrounding organs. The first step is to do a CT scan of the patient’s tumor and surrounding tissues. The scan will produce multiple cross-sectional images, or “slices,” which together effectively yield a 3D image.

“You take your medical imaging – there can be 100 images that are slices through the tumor – and hand-draw on each slice where the tumor is,” Mak explained. “Then you draw in where the organs are. And then you determine how to aim the radiation.”

That prep can take hours, Mak said. “It’s very time consuming,” he added. “And there’s lots of variation even between experts. When you miss and the radiation doesn’t go to the right place it can directly impact patient care.”

When you look at the issue from a global perspective, there’s also the problem of lack of access, Mak said. Many places do not have enough specialists to draw the tumors accurately.

To determine whether artificial intelligence could fill the void, Mak and his colleagues set up a contest on a website called that hosts challenges for a community of more than a million programmers around the world, who compete for prizes by solving computational problems.

In this case, the researchers had $50,000 in prize money to offer coders who could come up with “new AI techniques that could train machines to replicate an expert clinician’s ability to target a tumor,” Mak said.

A total of 564 contestants from 62 countries registered for the challenge, which was to design a program that would allow a computer to look at multiple expert-drawn tumors and thereby learn to draw itself.

In the end, 10 independent winning algorithms, filed by 9 participants, were chosen. Then the top five algorithms were combined into one “ensemble” program, which did fairly well when put up against the work of actual radiation oncologists.

“We compared the performance of the algorithms, that is, compared algorithm-generated (drawings) versus the human expert to generate a performance score on each case, and then benchmarked against the variation seen between multiple human experts against the study’s expert, and also the intra-observer variation, that is the same expert doing the same task twice,” Mak said. “The ensemble of the best algorithms had overlap scores in 75 percent of the cases that matched intra-observer score.”

The new approach may help fill the global shortage of radiation oncologists, said Dr. Sushil Beriwal, a professor of radiation oncology and deputy director of radiation services at the Hillman Cancer Center at the University of Pittsburgh Medical Center in Pennsylvania.

“One of the biggest challenges to targeting tumors is the lack of manpower,” said Beriwal, who was not involved in the new research. “The eventual goal is to come out with a product that can be used to help where an expert is not available and to use as a second check (of the radiation oncologist’s own work).”

In the final iteration of the project, “they were very close to the variation you would expect between physicians,” Beriwal said.

Dr. Nicholas Sanfilippo welcomed the new research. “It’s an exciting new technology,” said Sanfilippo, a radiation oncologist at NewYork-Presbyterian/Weill Cornell Medical Center in New York City, who also wasn’t involved in the study. “As clinicians we should look to making it complement the care of the physician, especially in underserved areas.”

SOURCE: and JAMA Oncology, online April 18, 2019.


Leave a Reply

This website uses cookies. By continuing to use this site, you accept our use of cookies.