Aaron Babier, a Ph.D. candidate at University of Toronto, and his team have developed a new method for AI to optimize radiation therapy plans. While AI has been used for optimization previously, Babier is taking a new approach.

The inspiration struck nearly two years ago. Babier said there was an optimization method that provided an output a clinician could use in their current treatment planning pipeline. “We knew that we could get those parameters if we had dose volume histogram curves, which are basically high-level representations of a treatment plan,” he said. “We did a bit more digging in the literature and found there were a lot of methods that would predict these high-level treatment plan features.”

After the team modified the features, they were able to get predictions for the target as well as the healthy tissues. Combining that information with their optimization method allowed for the creation of an optimization pipeline. “So the things we developed were based on the bag and query method and this generalized PCA (principal components analysis) approach, which was based on previous work,” Babier explained.

Babier was inspired to get into the field of cancer treatment after his stepmother passed away from cancer when he was 12. His interest in optimization and machine learning developed later, but ultimately led him to the idea of automated treatment planning.

Studying the field of automated planning, Babier realized that many of the current automated pipelines overhaul current planning paradigms. He realized there was an opportunity for a new method to be adopted if a higher level of control could be given back to clinicians since, he said, the current programs don’t allow clinicians to change the plans in an intuitive manner. “Our program can almost work like a black box behind the scenes of the current planning method and it can basically be a plug-in to a framework that clinicians are familiar with,” he said.

Essentially, Babier’s method isn’t looking to reinvent the wheel. Instead, it’s looking to add some rubber to the treads to help clinicians get more control of their plans in a way that provides support, but isn’t overbearing.

The team focused on throat cancer for their research because, as Babier explained, they wanted to study a cancer treatment plan that would be relatively complex, while having a heterogeneous patient population – or a cohort which looked very different from each other. The complexity comes in part from healthy tissue often overlapping the target area, “so it’s not as intuitive on how you should be planning for certain cases. We wanted to pick a site where the patients are very different so we could use the machine learning to predict the trade-off the clinician would normally make before the clinician looked at the plan. We chose head and neck because the trade-offs are quite variable between the patients,” Babier said.

Babier believes that by perfecting the software by focusing on head and neck, it should translate well to other sites like prostate and breast. The evidence backs him on this, as there have been successful implementations from head and neck to simpler sites in the past. He says the hard part is getting the data to do it.

Regarding the optimization, Babier says the work the software does takes about 20 minutes. For a human to do the same work – to go in and tweak optimization programs and then review – takes 15 to 20 minutes. However, it’s typically not a one-time tweak of a program. Babier says the process of adjusting, reviewing and revising might take a couple of hours to a couple of business days. So if it was “one and done” for a human to do the work, it would be one story, but due to repeat revisions the AI optimization earns its keep.

Babier believes there’s work to be done not just on the software and technical side, but also on the personnel side where the people working on machine learning and those in clinical practice aren’t necessarily speaking the same language, but both sides are starting to understand each other more as the work evolves.

One of the big takeaways from the work the team did was learning how messy hospital data is. Even within the treatment plans, Babier said there was a lot of information that wasn’t clear on how they were developed. “Even the simplest things, like how certain structures are named, can vary dramatically. Dealing with that weird naming culture is a challenge. The data cleaning work you do upfront is really important to these AI optimization processes. I will say that, apparently, hospitals are getting better with their labeling – I think they’re starting to use scripts or AI more to make it more uniform. When humans are in charge, variation happens very quickly,” Babier said.

The next step for the technology is to get it out to a hospital in Canada to further refine it, with the plan to grow it to other hospitals and other treatment planning software.