Researchers at Eindhoven University of Technology (TU/e) have developed a groundbreaking basic medical AI model that helps doctors recognise abnormalities on CT scans faster and more accurately. This enables earlier diagnoses of conditions such as cancer. The model was trained on more than 250,000 CT scans and is the first to use the computing power of SPIKE-1, TU/e's new supercomputer.
The AI model, developed under the leadership of Dr Fons van der Sommen, functions as a so-called foundation model: a basis on which hospitals, universities and companies worldwide can build to develop their own medical AI applications. Thanks to self-supervised learning, the model independently learns to make connections between images and accompanying texts, without researchers having to manually label each example.
Open innovation for healthcare
TU/e is making the model available as open source, so that institutions without a supercomputer can also use it. ‘We are, as it were, providing the seed from which others can grow their own medical AI models,’ says Van der Sommen, who is affiliated with the Architectures for Reliable Image Analysis (ARIA) research group and the Eindhoven Artificial Intelligence Systems Institute (EAISI). ‘In this way, we are lowering the threshold for innovation, especially in healthcare, where relevant data is often scarce, for example in the case of rare diseases.’
According to Van der Sommen, open collaboration is essential for progress in medical AI. ‘In the past, people preferred to keep such models to themselves,’ he says. ‘But this model can lay so many golden eggs that we cannot all handle it. By sharing them, everyone can move forward.’
The power of the SPIKE-1 supercomputer
Training such a large AI model was only possible thanks to the SPIKE-1 supercomputer, equipped with NVIDIA DGX B200 systems and the world's fastest Blackwell GPUs. With over 5.7 terabytes of internal memory, SPIKE-1 can process hundreds of CT scans simultaneously. This is something that is impossible for ordinary graphics cards.
‘A single CT scan can easily be a hundred megabytes. We had to process thousands of those images at once. Without SPIKE-1, that simply would not have been feasible,’ says Van der Sommen.
From laboratory to clinic
The researchers emphasise that AI is intended as a support, not a replacement for doctors. ‘We see it primarily as a qualitative tool for better data analysis. It takes a lot of work off our hands, especially in the area of specific signalling, but doctors are still needed to interpret those signals.’
The project has three objectives: to demonstrate what foundation models can mean for medical imaging, to find more efficient ways to train such models, and to strengthen TU/e's position as a pioneer in open-source AI research for healthcare.
In the coming period, the team wants to further share and expand the technology. Van der Sommen concludes: ‘We are laying the foundation, others can build on it. Universities have the knowledge and infrastructure to take the first step. Companies and hospitals can then convert that knowledge into applications that really make a difference for patients.’
AI for more accurate medical research
A few weeks ago, researchers at the University of Illinois Urbana-Champaign announced that they had developed an AI-driven tool that automatically detects missing or incomplete information in clinical trial reports.
The system analyses published articles for crucial elements, such as descriptions of randomisation, patient selection and outcome definitions. These are factors that are essential for the reliability of medical evidence. The research shows that many clinical studies are still reported incompletely, which makes it difficult to assess their quality.
The team is currently working on a lightweight, open-source version that will be made available to researchers, universities and medical journals worldwide. This will allow scientists to have their draft articles automatically checked for completeness, ensuring more reliable publications, faster quality control and greater transparency in medical research.
