Quietly, assuredly, relentlessly – RaySearch Laboratories is busy shaping a data-driven revolution in radiation oncology. The Stockholm-based oncology software company has, for the past decade and some, been making strategic investments in machine learning, automation and big-data technologies. Its vision: fundamental transformation of the radiation-therapy workflow – improving efficiency and consistency in treatment planning and delivery, while freeing up specialist clinical staff to dedicate more time to patient care.

If the vision is clear, so too is RaySearch’s laser focus on delivery against that vision. Back in April, at the annual congress of the European Society for Radiotherapy and Oncology (ESTRO 38) in Milan, two much-talked-about machine-learning innovations – automated organ segmentation and automated treatment planning – received top billing at the official European launch of RaySearch’s RayStation 8B treatment-planning software.

In combination, these new machine-learning applications represent core enabling technologies for the clinical implementation of online adaptive radiotherapy, with the end-game of personalized treatment plans (tailored to the unique needs of each patient) delivered in minutes rather than hours. “As a vendor, we can point to the efficiencies of automation and we can point to the consistency inherent to machine learning in radiation oncology,” says Fredrik Löfman, head of machine learning and algorithm at RaySearch. “But in terms of treatment quality and patient outcomes, it’s the clinics that will need to provide the real-world evaluation and validation.”

Machine learning visions

With that in mind, RayStation’s machine-learning capabilities are currently being road-tested and co-developed by a number of clinical partners, with more and more data emerging daily on the benefits for the radiation oncology workflow. Among those machine-learning early-adopters is the Laboratory of Artificial Intelligence in Radiation Oncology (LAIRO) at Massachusetts General Hospital (MGH) in Boston.

LAIRO’s team of eight clinical staff – three medical physicists and five medical dosimetrists – is headed up by Yi Wang, who prioritizes the clinical orientation of the laboratory’s research programme. “Our goal at LAIRO is to develop an intelligent platform for the whole radiation oncology workflow – segmentation, treatment planning, as well as outcomes and clinical decision support,” he explains. “We have this ‘explorer mindset’ to exploit emerging technologies – cloud computing, big data and machine learning among them – for greater workflow efficiency, lower healthcare costs and enhanced treatment outcomes.”

RaySearch is LAIRO’s main industry partner, and exclusively so on R&D for treatment planning. Right now, Wang and his colleagues are carrying out clinical studies using the machine-learning optimization (MLO) capabilities in RayStation to enhance treatment planning for liver stereotactic body radiotherapy (SBRT).

“We have provided a lot of clinical insight leading to the construction and integration of our MLO ‘auto plan’ models in RayStation,” explains Wang. “This latest study is a golden opportunity for us to build and test the first liver SBRT MLO model trained with multicriteria optimization [MCO] plans for RayStation.”

It helps, says Wang, that the liver SBRT is a well-defined problem. “With this treatment, we have many prior cases with a sufficient level of similarity and diversity to construct a good machine-learning model – simple enough that the machine can learn it, and meaningful enough that it can lead to improvements in workflow efficiency, also greater operator uniformity when the model is used by different treatment planners.”

The work has progressed to the point where the liver SBRT machine-learning model is now ready for prospective clinical evaluation. In parallel, the LAIRO team is working on several more challenging problems. For starters, there’s a machine-learning model to support radiotherapy treatment planning for pancreatic cancer addressed with simultaneous integrated boost (SIB) in the same fraction (i.e. a lower radiation dose to a larger treatment volume with a boost dose to a smaller volume, which can have variable size, shape and location within that bigger volume).

Other machine-learning models under development at LAIRO include lung SBRT with more complex tumour locations than the liver; SIB for head and neck cancers with more complex anatomy and dose pattern than the pancreas; and prostate cancer treated in sequential boost (i.e. a lower dose to a larger volume followed by a boost course to a smaller volume).

“We designed this roadmap with increasing complexity of anatomy and dose pattern,” explains Wang. “Ultimately these machine-learning models will help the treatment planners working across our network of clinical facilities, giving them a robust baseline and universal starting point based on prior clinical experience and data.”

It’s all about outcomes

The machine-learning applications in RayStation utilize models that have been trained on historical data – typically around 100 patients and plans are used for the training and validation. What’s more, deployment of the machine-learning models is independent from the version of the treatment-planning software – the models are effectively decoupled from RayStation – which creates unique opportunities for collaboration and knowledge-sharing between cancer centres.

For Wang and the LAIRO team, it’s workflow efficiencies and aggregate time-savings that are the biggest selling points of MLO. “With RayStation,” says Wang, “we can do one-click automated planning to generate clinically acceptable treatment plans with a quality very similar to the manual plans generated by our best treatment planners using multicriteria optimization.” In most cases, he adds, these auto plans are indistinguishable from manual plans when viewed by the attending oncologist.

Yet while MLO is all about statistical solutions based on prior clinical experience, there are also compelling opportunities to individualize radiation therapy so that that specific patients receive maximum benefit. Wang says this “multi-strategy” approach enables automated planning to create auto plans with different emphasis. One planning strategy, for example, might favour greater coverage of the target, while another might prioritize the sparing of healthy tissue.

“With one click, you can run different strategies and let the physician pick among them,” Wang explains. “It’s like you have four different planners creating four different plans, and the best auto plan can be further refined and personalized by post-processing. We’re heading towards personalized, precision medicine based on prior clinical knowledge and experience accumulated over decades. Machine learning is a great tool to make this happen.”

Continuous improvement

Meanwhile, Löfman and his RaySearch colleagues will continue to prioritize data-driven product innovation in tandem with clinical validation. That starts here and now by pushing machine-learning models beyond the quality of the historical treatment plans on which they are trained.

“It’s not just about capturing the best from the historical plans, it’s about continuous improvement going forward,” Löfman explains. “You don’t want to reach a point over time where machine learning is not progressing. For each new plan generated, machine learning needs to push as hard as possible – for example, in terms of reducing dose to organs at risk.”

Read more

Machine learning ready to shine in radiation therapy

Down the line, Löfman identifies data-driven treatments as the “next big thing” in radiation oncology. “This is where the whole field is heading – a greater emphasis on availability, accessibility and standardization of data to support optimized treatments and enhanced clinical outcomes. There are new roles for clinical staff here and a requirement for new infrastructure and greater collaboration between clinics.”

Equally significant for Wang is the opportunity to level the playing-field with a joined-up and networked approach to MLO. “In time, machine-learning models will be accessed by multiple clinics around the world – so that every clinic can start from the same baseline,” he concludes. “We are democratizing the accumulated knowledge and experience in radiation oncology and pushing that out across borders to a global user base.”