RaySearch Laboratories AB

Every second counts

RaySearch’s second AI application concerns the creation of treatment plans for individual patients, which involves finding optimal treatment delivery parameters to deliver the radiation. Whereas traditional solutions deal with this iteratively – setting up optimisation functions, generating solutions, and refining the optimisation until the plan is approved – RaySearch automates it by using deep learning to predict the dose distribution based on user prescriptions and automatically triggering an optimisation problem to get as close as possible to the predicted dose. “This automated process usually takes only a few minutes, with optimisation time being the bulk of it. The result is an automatically generated treatment plan ready for review, refinement, quality assurance and delivery, saving considerable time,” says Fredrik. “By combining the two applications, the process from image acquisition to automatically generating a deliverable treatment plan can be significantly streamlined, allowing clinicians to focus on reviewing and refining the end result before approval.”

“One notable aspect is that we integrate AI functionality directly into the code base, allowing for incredibly fast processing without needing to transfer data elsewhere. This not only provides a fascinating technical capability but also offers clinical benefits, particularly for adaptive treatments where patients are waiting on the treatment couch during computations. Every second counts, and our ability to process data quickly sets us apart. And this emphasis on speed is central to the ASSIST project’s focus on automation and AI.”

First of a kind

ASSIST marks RaySearch’s first ITEA project and the motivation for this came initially from the opportunity to build partnerships with hospitals, universities, and other companies. But, as they dug into its details on optimising and simplifying image-guided therapy workflows, two aspects came to the fore: federated learning and automated workflows that use just magnetic resonance imaging (MRI). For the latter, they teamed up with fellow Swedish company Spectronic Medical to combine their software capabilities to achieve an MRI-only flow, thereby skipping the standard, time-consuming computed tomography (CT) scan.

Federated learning, meanwhile, is particularly relevant to industries with regulations and legal aspects that make it challenging to access data. Instead of bringing data to the AI algorithm, the AI algorithm goes to the data. Within ASSIST, RaySearch conducted a pilot to train their dose prediction model by splitting a dataset into different nodes and sending the algorithm to these nodes before combining the results into a model. “Now, we are involving Swedish university hospitals in this exercise, making it the first time a dose prediction model has been trained this way,” notes Fredrik. “Hospitals publish data for a certain treatment protocol, and we use our dose prediction algorithm to train on these nodes without the data leaving the hospital, ensuring privacy and regulatory compliance. Training on data from different hospitals also makes the model more robust than training on data from a single source.”

Positive partnerships

In addition to the work in their own field, Fredrik values the project’s insights into other areas of healthcare, including hardware to get a better view of patient geometry and the use of automation and AI in surgery. And although they were already accustomed to working with clinics in Europe, North America, and Asia-Pacific, involvement in ITEA added a new dimension to their work through collaboration with international companies and universities – something that Fredrik expects to have long-lasting benefits even after the project comes to a close in September 2024.

“The project management team, led by Philips, has done a superb job in driving this project efficiently and transparently despite involving a complex ecosystem of participants. It’s been a valuable experience witnessing such effective project management. We’ve also engaged in discussions with clinics and companies outside of Sweden, including hospitals that we have collaborated with before. Additionally, before this project, we had only limited contact with Linköping University, which is driving the Swedish part of the consortium. Establishing this connection has been very positive, and we anticipate continued collaboration with them after the project ends.”