Machine learning technology helps small clinic maintain quality and consistency while meeting the growing demand for radiotherapy services.
Since its deployment in 2013, RapidPlan® knowledge-based treatment planning software has been widely deployed at many cancer centers around the world. Still, a perception exists among some clinicians that RapidPlan is useful mainly for large institutions managing massive patient volumes, or for achieving consistency in treatment plan quality across a disbursed network of sites.
In this article, “Centerline” profiles a smaller treatment center, the regional hospital of Gävle, which serves two counties in northern Sweden, Gävleborg and Dalarna, with a combined population of around half a million people. Each year, the hospital treats about 1,250 cancer patients, mainly for breast, prostate, lung, head and neck, and brain cancers, and for palliative care.
The Gävle regional hospital began using RapidPlan in 2017. Right from the start, medical physicists Per Hållström and Nils Andrae were pleasantly surprised at how easy it was to implement, even for a small clinic.
“Early on, our goal was to use RapidPlan for head and neck sites, but we decided to start with some prostate salvage cases that were less complex,” explained Hållström. “Planning for head and neck is a very complicated process and can involve many iterations for quality control, depending on the skills and experience of the treatment planner. We wanted consistent, high-quality plans to give our patients the best possible treatment.”
To meet regional requirements for target and organ-at-risk delineation, the Gävle team built its own RapidPlan models based on high-quality prostate salvage treatment plans they had already created in the process of treating patients. As the medical physicists observed the model’s effectiveness, they moved to head and neck planning with RapidPlan the following year. So far, the clinic has treated about 85 head and neck cancer patients and 460 prostate cancer patients with plans produced by RapidPlan.
“We already had high quality prostate treatment plans to use in creating RapidPlan models for prostate, but needed at least 20 to 25 head and neck patient plans for creating that model, not including outliers,” said Andrae. “Of course, the more patient plans you use to train the model, the better, and it’s important to compare plans against additional test patients to verify the results.”
Over the past three years, the team has made three revisions to the RapidPlan head and neck model based on 45 patient plans, and the prostate model now has on 108 plans in its database.
“Obviously, we don’t treat as many head and neck cases as prostate, but we’re seeing excellent plans from RapidPlan,” Hållström said. “Even if the tumor is quite deformed, the plans are very good and when we optimize, only require minor tweaks.”
The Gävle cancer care clinic has also shared its models with other sites, including the Karolinska Institute in Stockholm, which started using RapidPlan slightly later, and according to Hållström also builds its own models.
“We follow the same guidelines, even though different doctors may have certain preferences,” Hållström said. “The feedback on our models was that they got good results, and that’s one of the benefits of sharing RapidPlan models. You can see how others have set up the models and use their comparisons to help you get started.”
With both prostate and head and neck models well established in the system, the Gävle team is seeing the consistent, high-quality plans they had initially hoped to achieve.
“QA has always been an important part of our process, comparing plans with the manually optimized results of older patient plans, and using Mobius3D for review,” Andrae added. “For prostate, we typically accept the first plan straightaway and for head and neck, we may have to do two iterations at the most. A huge benefit has been that planning time has been reduced from hours down to about 30 minutes.”
Furthermore, Andrae explained, advanced features in RapidPlan mean they no longer have to use support structures and crop them out later, and all organs at risk (OARs) are accounted for in the first plan.
“You always have everything you need in the initial plan, which saves time and gives us confidence that we’re making the treatment as safe as possible for the patient,” he said. “Even if the doctor decides to change something related to the target, once you have the model it’s easy to create a new plan.”
Despite being a small clinic, both medical physicists agree that the initial time and effort they put into developing the RapidPlan models have been well worth the time savings and reassurance that they’re offering their patients the best possible treatment plans.
“Right from the beginning, the initial planning results are always very good—we’ve never had to scrap a plan and start again,” Andrae explained. “We measure every single plan on the machine and over time we’re seeing them improve more and more. We certainly would not want to go back to the days when we didn’t have RapidPlan—we definitely don’t miss those days.”
Photo caption: The team in Gävle, from left: Sandra Englund Freimuth, Anna Hellman, Birgitta Persson, and Towe Lindell (each a nurse and dose planner); Per Hållström and Nils Andrae (medical physicists).
Further reading:
RapidPlan Machine Learning Brings Intelligence to Cancer Care Globally (December 6, 2019)
Machine Learning at Northwell Health: Using RapidPlan to Develop Quality Head & Neck Treatment Plan Models (June 4, 2019)
Research Demonstrates RapidPlan Treatment Planning Can Improve Plan Quality (July 12, 2017)
RapidPlan Knowledge-Based Planning: Clinical Adoption Takes Off (June 7, 2017)
The information captured herein represents the genuine experience of the attributed individuals and may not necessarily represent the views of Varian or the above-referenced institution. Individuals were not compensated for their participation. Radiation treatment may not be appropriate for all cancers. Individual results may vary. For more information, please visit www.varian.com/safety.