AI server for applications in medical physics

• Subject Area: Particles, Nuclei and Fields ; Condensed Matter Physics
• Term: Funded in 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 569432750

Medical physics is a prime example of a research area where large amounts of data are collected and analyzed. The data comes from a variety of sources and includes 2D and 3D images from different imaging modalities, such as magnetic resonance imaging or computed tomography, as well as data from 3D dose predictions and simulations. Typical tasks include regression and classification, as well as optimization and numerical calculations. In many cases, these tasks are solved using AI methods. Due to the large amount of data and the high complexity of the methods used, a correspondingly large amount of computing power is required to solve these tasks successfully. Experimental particle physics and medical physics, in particular radiotherapy, face similar challenges in data analysis. In particle physics, data often comes from large-scale experiments with millions of readout channels and high data throughput. They also come from the simulation of such experiments and the calculation of theoretical predictions. While the energies of the particles and the scale of the experimental setups are significantly different in these two fields, the tasks involved in analyzing the data are largely the same. The resulting great potential for synergies between the two areas, for example in the development of surrogate models, classification problems and other statistical approaches, will be exploited. The AI server to be procured consists of multiple graphical processing units (GPU) with Tensor Core technology connected to ultra-fast switches and large Video Random Access Memory (VRAM). Such a configuration is key to training modern AI algorithms with complex architectures and reasonable batch sizes. It enables the study of hyper-parameters of the algorithms, resulting in optimal performance of the models and the data used. In medical physics, the AI server will be used, for example, to develop surrogate models in simulations for dose predictions, to calculate dose distributions on MRI images and to process (complex-valued) three-dimensional image data. While the training of high-precision models is known to be resource-intensive, their predictions are easily calculated with low hardware and energy requirements, even in complex applications. This can be crucial for real-live applications, for example, to enable online treatment adaptations in radiation therapy. Application examples in particle physics with high synergy potential include the classification of different process classes, the sampling of high-dimensional spaces and the improvement of Monte Carlo generators. The server will exploit the full potential of the methods and data so as to gain new in-sights into medical physics applications and answer selected questions in particle physics.

• DFG Programme: Major Research Instrumentation
• Major Instrumentation: KI-Server für Anwendungen in der Medizinphysik
• Instrumentation Group: 7030 Dedizierte, dezentrale Rechenanlagen, Prozeßrechner
• Applicant Institution: Technische Universität Dortmund
• Leader: Professor Dr. Kevin Kröninger