The model was trained on more than a quarter of a million CT scans. The research team was the first to use the computing power of the new SPIKE-1 supercomputer for this purpose. TU/e is now making the AI model available worldwide to universities, hospitals, and companies, allowing them to further develop it for their own medical applications. This is especially useful for institutions that lack access to a supercomputer of their own.
The team, led by Associate Professor Fons van der Sommen , trained a so-called foundation model: a baseline system upon which other, more specialized AI applications can be built.
The researchers applied self-supervised learning, a method in which the model independently learns to identify relationships between images and text, without requiring humans to label every single example manually. Testing and refining the method's usefulness formed part of the research.
The model was trained using approximately 250,000 three-dimensional CT scans and 75,000 radiology reports from public medical databases. These included various disease types, including cancers. Sometimes the scans showed no abnormalities, which was equally valuable, as healthy images also help improve model accuracy.
Model freely available for global use
"You can think of it like grafting a plant," says Van der Sommen, who is affiliated with the research group Architectures for Reliable Image Analysis (ARIA) and the Eindhoven Artificial Intelligence Systems Institute ( EAISI ). "We provide the stem from which others can grow their own medical AI models. This lowers the threshold for innovation and collaboration in healthcare, because not everyone has access to the computational power and capacity that we do."
Hospitals often lack sufficient data as well. "The more data you have, the better you can train an AI model," Van der Sommen explains. "But in healthcare, the right data isn't always available, for example, in the case of rare tumors. As a result, we need to wait a considerable amount of time before meaningful breakthroughs are possible. With our model, we can now make a real difference for certain rare diseases."
We provide the stem from which others can grow their own medical AI models. This lowers the threshold for innovation and collaboration in healthcare.
Associate Professor Fons van der Sommen
The model will be available open source, enabling hospitals, research institutions, and companies to develop their own tailored variants. It can serve as the basis for AI systems that detect tumors, predict disease progression, or identify other patterns in medical imaging data.
Van der Sommen: "Previously, a new AI model was viewed as a goose that lays golden eggs - something you wouldn't want to share. But this model can lay so many golden eggs that we cannot possibly handle them all ourselves. By sharing it, everyone can move forward."
The supercomputer that made it possible
The research was only possible thanks to the computing power of SPIKE-1, the new supercomputer put into service by TU/e last year . It consists of four NVIDIA DGX B200 systems, each equipped with eight powerful Blackwell GPUs - currently the fastest AI chips in the world, designed for demanding tasks like training large models.
With more than 5.7 terabytes of internal memory, SPIKE-1 can process hundreds of CT scans simultaneously, something standard graphics cards simply cannot handle. "A single CT scan is easily around 100 megabytes," Van der Sommen explains. "For this research, we had to combine thousands of those images into a single training pipeline. Without SPIKE-1, achieving this result would have been completely out of reach."
AI to assist, doctors to decide
Van der Sommen continues: "In particular, Cris Claessens and Christiaan Viviers, two researchers from my group, played an important role. And the support staff at EAISI, TU/e's AI institute, contributed significantly as well." The project marked the first practical deployment of SPIKE-1. The researchers helped configure the hardware and software to ensure the supercomputer delivered optimal performance from day one. It operates in a sustainable data center in Finland.
AI can take over a lot of work, but doctors remain essential for interpreting the signals.
Associate Professor Fons van der Sommen
The project marked the first practical deployment of SPIKE-1. The researchers helped configure the hardware and software to ensure the supercomputer delivered optimal performance from day one. It operates in a sustainable data center in Finland.
Impressive as it is, the model will not replace human expertise. "We see it primarily as a high-quality tool to enhance data analysis. It can take over some of the more labor-intensive detection work, but doctors remain essential for interpreting those signals."
Three-fold objective
According to Van der Sommen, the project has three clear aims. "We want to demonstrate what foundation models can achieve in large-scale image analysis. At the same time, we sought the most efficient way to build such models. And thirdly, not altogether unimportant, we aim to strengthen TU/e's position as a leader in open-source AI research for the medical sector."
The researchers intend to continue publishing their findings. "This will help increase awareness of the results, and thus the international visibility of both the research and the model we developed."
From research to real-world impact
While the current focus is on medical imaging, the underlying technology has broader applications. Van der Sommen hopes the insights gained here will lead to new products and spin-offs. "We are paving the way toward clinical deployment," he concludes. "Universities have the expertise and infrastructure to take that first, difficult step: building and validating reliable models. Most companies and smaller organizations lack the means to do that, but what they can do, is bring the results further into society."
