Coding against cancer

As we step into the iLab, we can feel his passion. “I work with so many ambitious people from various countries here. Switzerland is an exciting and inspiring place to live and work,” Strzelecki shares. Originally from Poland, he spent several years working in France before making the move to Switzerland. How did he end up here? “I was always interested in computer graphics and also the technology around image processing or image generation.” After exploring 3D rendering for gaming and satellite imaging, Strzelecki was introduced to the field of medical imaging through a project during his PhD studies. “I discovered there is a lot of potential and wanted to continue exploring that, so I came here,” he says.

“If the image is no longer accurate on treatment day, you won’t precisely hit your target,” Strzelecki explains. “You might eradicate the tumor, but you could also harm the patient, because you just destroyed that organ at risk – that is vital.” Strzelecki and his team wanted to improve precision. The result is the HyperSight imaging solution – an optional feature on Varian’s Halcyon, Ethos, TrueBeam, and Edge radiotherapy systems with better image quality, precision, and speed.¹

But what is the difference between simulating on a treatment device and computed tomography (CT)? Traditional CT scanners use fan-beam and a strip detector to capture quick images of a strip of the body as the patient moves through the scanner, whereas LINAC systems use a technology called cone-beam CT (CBCT). The beam is shaped in the form of a cone and a large rectangular detector captures a series of images from different angles. Combined with sophisticated mathematics, the images create a 3D representation of the body – a digital twin of the patient. This allows for an accurate determination of distances and locations of both the tumor and surrounding healthy tissue, crucial for precision therapy.

"What makes imaging with CBCT challenging is that it captures the entire scan in a single arc. Any movement during the scan can significantly affect the image quality and lead to artifacts. The cone-shaped beam results in higher scattered radiation. This makes images blurry, and it is difficult to see fine tissue structures," Strzelecki explains. Together with his colleagues he took on the challenge to improve image quality of CBCT to use the images for treatment planning.

Strzelecki coordinates technical tasks and fosters an environment where they all work together to solve problems: “I took on the role of being aware of the overall situation and managing priorities. That includes current and new platform developments, customer feedback, and maintenance releases.”

What he enjoys most about his work? “That I can learn. I am a curious person. I see something and I want to understand how it works,” says Strzelecki. His fascination with technology was influenced by his father, a mathematician who always encouraged him to seek out answers independently. His free time is no exception to that: “My hobbies change a lot, but they often involve learning nerdy things like fixing stuff myself or reading about retro computers.”

HyperSight is a collaborative effort of different research and development teams. While Strzelecki’s team focused on image reconstruction algorithms, the software team worked on a new user interface for cone-beam CT to test the new workflow for planning with user experience (UX) studies. The hardware team developed bigger imaging panels that can produce larger images. For example, the new imaging panels on the Halcyon and Ethos radiotherapy system allow for a shorter 211-degree rotation instead of a full 360-degree rotation which reduces acquisition times by a factor of two.

All this work was done simultaneously. “The biggest challenge was developing the new algorithms without having the new hardware,” he says. “We had to work with virtual models most of the time. When we got access to it later, we had to creatively address problems, as there wasn’t much time to explore multiple theories.”

Strzelecki appreciates working in a team with different professional backgrounds, skills, and perspectives. It allows them to support each other and overcome challenges. “We are very complementary,” he says. “Our culture supports bringing all these different people together that bring a variety of ideas to the table. It is a great experience to create something together.” He believes in the importance of sharing a common mindset and motivation: “I hope we can attract new brains that can support our vision and my team’s work.”

“We now have image quality comparable to planning-CT on cone-beam CT,” Strzelecki states. “Especially for patients with tumors in sensitive or difficult-to-view areas such as the pelvic region or the lungs, therapy can be more precise and personalized.” In our adaptive therapy platform, once the image is available, AI detects different structures and updates the treatment plan using AI-based segmentation. This can lead to more effective treatments and better patient outcomes. 

Faster imaging can shorten treatment time. “Patients have to hold their breath up to 60 seconds in traditional CBCT scans. This can be uncomfortable and trigger anxiety. With our improvements we can minimize this as patients spend less time on the treatment table and only need to hold their breath briefly,” Strzelecki explains. Delays during treatment can be avoided as there is no need for additional trips to a separate CT scanner. “When we create clinical workflows that are more efficient, that allows more patients to be treated,” Strzelecki says.

Strzelecki’s team collaborates routinely with clinical partners to further improve image quality and ensure their solutions are reliable in real-world settings. “Phantoms in labs cannot replicate the complexity of real patients. Getting clinical data and feedback from clinical partners is crucial for identifying problems. It ensures our imaging solutions represent clinical features accurately on patient images,” Strzelecki explains. They also collect a wide range of clinical cases and refine the algorithm to work across different scenarios.

“Our highest priority right now is developing the next generation of the CBCT reconstructor,” Strzelecki says. “The existing architecture has its limits, so we are creating a new GPU-centric algorithmic stack that can handle more complexity.” The goal is to simplify the codebase, making it easier to drive future developments. They are also working on using machine learning and AI for imaging: “This requires strong control to ensure accuracy. Future releases will likely be hybrid solutions, combining AI with conventional methods to verify the agreement between AI outputs and real measurements.”