Improving dose prediction in microbeam radiation therapy using conditional flow-based neural networks

• Applicant: Dr. Marco Menen
• Subject Area: Medical Physics, Biomedical Technology ; Methods in Artificial Intelligence and Machine Learning
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 580386328

In recent years, much progress has been made towards optimizing cancer treatments with radiotherapy, as new treatment methods based on different beam setups have emerged. A promising example is microbeam radiation therapy (MRT), which uses a grid of micrometer-wide X-ray beams and has shown reduced damage to healthy tissue compared to conventional radiotherapies. Accurate dose predictions for MRT, however, remain challenging as no patient data is yet available and the required high resolution makes Monte Carlo (MC) simulations computationally very expensive. Generative neural networks (NNs) have been shown to speed up dose predictions while retaining good accuracy, making them well-suited for this application. The goal of this project is to understand how a conditional flow-matching (CFM) network can be used to improve dose predictions for MRT, how it compares to other NN architectures, and what features of the simulation it exploits for data sampling. The CaloDream model, which has previously achieved state-of-the-art accuracy in calorimeter shower generation in particle physics, will be employed. It will be trained for efficient MRT dose prediction and benchmarked against previously used NN architectures. The results will be integrated into the open-source treatment planning system matRad. Finally, symbolic regression will be applied to extract features of the learned CFM network and improve the interpretability of the results.

• DFG Programme: WBP Position