Machine learning model creates treatment plans for patients with prostate cancer

August 30, 2021

Health Insights Machine Learning Success Stories Trustworthy AI

By Ian Gormely 

August 30, 2021

A new machine learning model can create radiation therapy treatment plans for patients with prostate cancer. The model, which produces plans deemed as good or better than human-created plans about nine times out of 10, reduces a process that can take more than a day to a matter of hours, freeing up valuable hospital resources. 

The team behind the model, which includes Vector Faculty Affiliate and UHN researcher Chris McIntosh, detailed their work in the paper “Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer,” which made the cover of Nature Medicine in June. It’s believed to be the first model of its kind and is currently in use at Princess Margaret Cancer Centre in Toronto.

“About 40 percent of patients with prostate cancer receive radiation therapy,” says McIntosh. “Traditionally, creating a treatment plan is a complex, iterative process specifying where and how radiation is delivered to the patient. We set out to automate that process using computer vision.”

The team honed the model for several years before deploying it for prostate cancer patients at Princess Margaret. Clinicians were presented with two treatment plans: one AI-created and one human-created. In a blind test, both were evaluated for their quality of care with a clinician making the final decision about which to initiate. They also ran a modified Turing Test, asking clinicians which plan they thought was generated by an AI. “That was pretty crucial,” says McIntosh. “We could go back and judge how often clinicians were correct and check for clinical bias for or against AI.”

While the model’s plans were “clinically acceptable” 86 percent of the time, they were only selected for patient treatment 60 percent of the time. “There was a hesitancy to accept the AI’s plan,” says McIntosh. “Clinicians get more hesitant as lives are at stake. Some skepticism helps ensure patient safety, but If the technology continues to prove itself overtime, it will abide, I think.” 

Prostate cancer was targeted first both because of how common it is (fourth in Canada) and the enthusiasm they received from the team in the clinic. But the model can be expanded to any radiation therapy for cancer. “Bandwidth is the primary thing holding us back.” 

In order to build a clinically deployable system, the team partnered with RaySearch Laboratories, a world leader in radiation treatment planning software and the provider of Princess Margaret’s pre-existing clinical treatment planning system. The partnership, in which their tech was licensed to RaySearch, also gives McIntosh and his team the ability to help cancer patients globally. Princess Margaret is one of the top five cancer centres in the world. Says McIntosh: “This technology means we can export that expertise and help more patients.” 

Check out Vector’s health page for more information about our health AI activities. 

Related:

Vector Faculty Member Gautam Kamath
Insights
Research

Vector researcher Gautam Kamath breaks down the latest developments in robustness and privacy

Research
Trustworthy AI

World-leading AI Trust and Safety Experts Publish Major Paper on Managing AI Risks in the journal Science

Insights
Large Language Models

Standardized protocols are key to the responsible deployment of language models