Machine learning drives advances in automated treatment planning

By Sharon Breske


Artificial intelligence (AI), machine learning (ML), organ segmentation—they may sound like the makings of a medical sci-fi movie, but in radiation treatment planning, that fiction is reality and the future is now.

A case in point is the RayStation 8B, the first treatment planning system to incorporate ML applications, according to the Swedish med-tech company RaySearch. The company debuted the system in December 2018 and recently released several updates in the RayStation 9A.

RayStation combines adaptive therapy, multicriteria optimization, and robust algorithms for intensity-modulated radiation therapy and volumetric-modulated arc therapy (VMAT) optimization with highly accurate dose engines for photon, electron, proton and carbon ion therapy. The system creates a high-quality treatment plan within minutes, bypassing the need for user intervention. With FDA 510(k) clearance awarded in July for deep learning (DL) organ segmentation and ML for automated planning for a key model, RayStation supports a wide range of treatment machines, providing one control center to synthesize treatment planning across the healthcare industry.

Model Sharing

With the high generalizability of segmentation algorithms, the clinical team can replace manual contouring with automatic contouring tools for organs on a particular model. “Then [the users] can fine tune them if they would like, or just go on and approve them,” says Fredrik Lofman, head of machine learning at RaySearch. “The deep neural networks have no patient data in them, so they are perfect to share. That means that you can plug a model into another RayStation at another clinic, in another country, or continent and produce identical contours.

“We don't have to change a single thing in the model training or model algorithm because it [adapts] to all the sites,” he adds, noting that planning models include prostate, head and neck, pancreas, liver, and lung, with breast models on the way, he notes. “Every model needs to be tuned for a specific protocol and the needs of the clinic. But you only tune the model once then you apply it to all the patients instead of [the method used] today where you create treatment plans and tune the optimization for each patient. That's a big difference.”

Among benefits is time savings. But that’s not due to the short time required to generate a RayStation treatment plan (about 40 to 50 seconds, sometimes lower), clarifies Lofman. It’s due to time saved in contouring the plans. “The time you want to minimize is the time to approve the contours,” he says. “This can drive us—especially if you look at multiclinic and multicountry models—toward a more consistent way of contouring. It's efficient because it's fast. But in the long run, it's really [about] consistency because … you will always get the same contours at the clinic.”

As with many AI-driven innovations, such capabilities prompt questions and concerns, generating healthy debate about automation. “What about the bias of the models? What about my job then? What about the validation of the models? And will [physician] knowledge of how to contour, for example, degrade over time if this is automated?” poses Lofman. “These discussions are really important.”

Regarding bias, Lofman says automated models that are trained and validated in a cross-clinic manner help mitigate this concern. But clinicians harbor biases, too. “If you're afraid of bias in models, I would also be afraid of bias in humans doing the contours. I would rather have the consensus of more than one person.”

New Features with RayStation 9A

Looking ahead, several planning models are undergoing validation for future FDA 510(k) clearance. The latest release, RayStation 9A, supports treatment machines with a dual-layer multileaf collimator and nonrotating jaws, properties of the Varian Halcyon system. These updates support RaySearch’s mission to unify treatment planning for as many systems as possible.

Also new is support for multicriteria optimization for VMAT in combination with Varian machines, a functionality introduced for Elekta machines in RayStation 8B. Additional possibilities for exporting VSim plans to other treatment planning systems are also available, as are new features for proton therapy planning, including caching of spot doses to accelerate optimization. Improved machine modeling for photons is now provided as well, allowing diagonal profiles in the beam-commissioning module of RayPhysics.

For carbon ion planning, support for Toshiba machines is now available too. Included is the microdosimetric kinetic model for calculating radiobiological equivalent dose, which helps support carbon ion treatments in Japan, home to most carbon ion facilities. For Siemens carbon ion machines, support for connectivity with the IONTRIS interface has also been added.

General improvements include support for a 6D couch during planning, and integration enhancements with RayCare—RaySearch’s oncology information system technology—such as automatically generated treatment plans based on a patient’s diagnosis.

Under Study: ML for Prostate Cancer

In May 2019, the first patients were treated using ML-generated RayStation treatment plans at the Princess Margaret Cancer Centre in Toronto. Patients with localized prostate cancer are receiving the treatment as part of a compensative evaluation study.

The prospective trial, headed by radiation oncologist Alejandro Berlin, MD, MSc, was launched after clinical results from a 2018 retrospective evaluation study found that ML plans were preferred or deemed equivalent to previous manual plans in 94% of cases. The results were based on three blinded expert reviewers.

The ongoing phase of the study presents physicians with two blinded treatment plans: a manually generated plan and a machine learning plan. The selected plan undergoes standard peer review and quality assurance, followed by treatment delivery with the preferred plan.

Ultimately, the study will provide unique data to quantify real-world performance and preferability of ML plans. “To date, [results] validate our observations about the robustness of this planning solution,” said Dr. Berlin in a company news release. “It has been really exciting for the team to help materialize this machine learning advancement in the radiation oncology field, including deployment into the clinical realm.”

Back To Top

Machine learning drives advances in automated treatment planning.  Appl Rad Oncol. 

January 21, 2020
Categories:  Technologies|Section

Copyright © Anderson Publishing, Ltd

    Agility CMS