Language Select

Unleashing the Power of Proton Therapy with Machine Learning

By Christel Smith, PhD MBA, Product Manager at Varian Medical Systems

Proton therapy is a rapidly developing cancer treatment modality with great potential to increase quality of life during and after cancer treatment. One advantage of proton therapy with pencil-beam scanning is that it offers improved accuracy and reduced dose to healthy tissues. If we look at proton therapy centers that are currently under construction or close to breaking ground, the number of operational treatment rooms will likely double within the next five years.¹

While patients will have more access to proton therapy, the growth of new centers presents new challenges. Particularly, as more proton therapy centers come online, the number of clinical professionals with previous proton therapy experience may be limited. How do we bridge the clinical treatment planning and management learning curve? How do we ensure that the quality of treatment plans does not vary across institutions due to the experience and knowledge of each individual planner? Additionally, how do we efficiently and effectively decide which patients would benefit from proton therapy instead of traditional radiotherapy?

One way to mitigate these challenges is to mine the libraries of previous patients treated with proton therapy to extract knowledge. This knowledge could then be distributed amongst the growing proton network to help bridge learning curves and provide a quality assurance framework. It may sound abstract, but Varian’s RapidPlan™ Software has already demonstrated that a knowledge-based approach to treatment planning in radiation therapy can increase quality while reducing inter-planner/clinic variation. With the current development of RapidPlan for protons, we are looking forward to bringing this technology to the proton community.

How it Works
In a radiation oncology context, machine learning works by training a model based on a set of previous patients using anatomy information, beam geometry, and a representative measure of the treatment plan quality called the dose volume histogram (DVH) as inputs. The model can then predict the DVH for a new patient, representative of the clinical practice captured in the scope of the model and clinical goals. The predicted plan for a new patient can be quite useful in the treatment planning process as it provides the treatment planner an informed optimization starting point, greatly reducing the need for time-consuming, manual trial-and-error processes.  

Several recent studies support the benefits of applying a knowledge-based approach to treatment planning. In a photon case study comparing the knowledge based planning (KBP) approach to manual plan generation at Radiation Oncology Queensland, Australia, (ROQ), manual input took 3.5 times longer compared to using RapidPlan.² Also in this study, investigators found that RapidPlan closed the quality gap between unexperienced and experienced planners--both groups were able to generate high quality plans using RapidPlan regardless of their level of expertise. Another study from the University of Washington, St. Louis reinforced this result.  By examining plan quality variation, they found that the KBP approach significantly reduced variation among treatment planners while increasing the quality of plans.³

Knowledge-Guided Decision Support
As an increasing number of patients gain access to proton therapy, how can clinicians determine which patients would be appropriate to refer to a proton therapy center for treatment? Currently, the decision to treat patients by protons or photons is done by generating comparative treatment plans, which is time-consuming and challenging. In the first study of its kind, researchers from VU University Medical Center (VUMC) in Amsterdam, were able to build a decision support tool using RapidPlan predictions. They did this by creating radiotherapy and proton therapy RapidPlan dose prediction models for head and neck cancer and then used these models to predict which treatment modality would spare more healthy tissue. According to VUMC researchers, RapidPlan model-based dose prediction capabilities may help eliminate the need to make a comparative plan, allowing for quick decisions based on a chosen threshold:

“This is the first investigation which demonstrates the feasibility of patient selection for proton therapy based solely on patient-specific knowledge-based predictions of proton and photon plan dosimetry, without necessitating actual plan creation.”⁴

In the near future, this knowledge-based planning approach could also potentially be used to predict the side effects of proton therapy vs. radiotherapy. By linking the normal tissue complication probability models to each treatment plan model for each modality, clinicians could predict the likelihood of a patient-specific adverse side effects based on different estimated doses. This approach could easily be extended to take patients’ preferences into account, allowing them to choose their treatment based on likelihood of toxicities.

As advancements in physics, radiobiology, and software technologies continue to accelerate, we are just starting to explore the full potential that proton therapy may hold. Solutions like Varian’s RapidPlan software can help clinicians unleash the power of proton therapy by helping to ensure the highest quality of care while also paving a much-needed pathway for decision support. Through continued work with our clinical partners, we are hopeful that we can continue to find innovative solutions that address real clinical needs with the goal of making a difference in cancer care.

1MedRaysIntel Proton World Therapy Market Report (2016 Edition)

2Case Study: Assuring Quality and Consistency in Treatment Planning with RapidPlan at Radiation Oncology Queensland 

3Moore, K.L. et al. Experience-Based Quality Control of Clinical Intensity-Modulated Radiotherapy Planning. IJROBP 81(2): 545-551 (2011)

4Using a knowledge-based planning solution to select patients for proton therapy.”  Delaney, A. et al. Radiation Oncology (in press) 2017.