Artificial intelligence (AI) can be helpful in analyzing medical images and support radiotherapy planning workflows. However, huge amounts of sensitive patient data are needed to train robust AI models. Researchers at the Karlsruhe Institute of Technology (KIT) have investigated how hospitals can jointly develop AI models for radiation therapy applications without exchanging such data. In their project, they use federated learning, a process in which a single model is trained at a decentralized location while the data remains in place. Their results demonstrate not only the technical feasibility of the approach but also its safe, more sustainable, and accessible use.
The exchange of medical data between hospitals is often limited by privacy regulations and organizational barriers, which hinders the development of robust AI models. A special feature of federated learning is that the patient data never leaves the institution where it was recorded. The facilities involved train their models on-site, exchanging only information about the model updates. This enables knowledge from multiple hospitals to be pooled without transmitting sensitive health data or storing it at a central location. While the approach is not new in the health sector, KIT researchers have refined it specifically for applications in radiation therapy.
Radiation therapy planning requires the use of various imaging techniques, such as MRI, CT, or CBCT. Each of these techniques provides different information. While CT scans are required for calculating the radiation dose, MRI provides high-contrast visualizations of organ tissue and CBCT scans enable precise positioning of patients during treatment. AI models can help synthesize CT images from existing scans.
"In the future, these images could render additional CT scans unnecessary in certain use cases and thus reduce additional ionising radiation and costs. This also eliminates the complex alignment of different image data sets, which can introduce inaccuracies," said Ciro Benito Raggio, a researcher of Professor Maria Francesca Spadea's Biomedical Imaging for Modeling and Simulation research group at KIT's Institute of Biomedical Engineering.
Technical Feasibility Demonstrated in Multiple Hospitals
The researchers developed and tested federated learning techniques for synthesizing CT images of the brain and the head and neck region. The results demonstrate that such AI models can be trained across multiple hospitals without the need to exchange sensitive patient data. The researchers consider this an important step toward clinically usable AI applications for radiation therapy.
Considering Security from the Outset
In addition, the research group studied potential security risks involved with federated AI systems. For example, they investigated whether image data from individual patients can be reconstructed from a trained AI model. The researchers demonstrated that comparatively minor adaptations are sufficient to secure federated AI systems against attacks on data security. "Data protection alone is not enough. If federated systems are to be used in medical practice in the future, we have to take potential security risks into account right from the start" said Raggio.
Lower Computing Costs and CO₂ Emissions
Besides data protection and security, the researchers also concerned themselves with the sustainability of theproposed federated learning solution. "Depending on the model architecture, we were able to reduce the time required for the training and CO₂ emissions by 9 to 23 percent without affecting the quality of the synthetic images. Smaller hospitals in particular could benefit from such optimizations, as they often don't have a lot of computing resources" said Raggio.
The results demonstrate that federated learning has the potential to support the development of privacy-friendly AI applications for radiation therapy while also providing access to powerful models for smaller facilities. The technology could thus contribute to achieving a broader integration of innovative approaches into clinical practice and medical care.
Original Publications
Raggio, C.B. et al. FedSynthCT-Brain: A federated learning framework for multi-institutional brain MRI-to-CT synthesis. Comput Biol Med 192(A), 110160 (2025). https://doi.org/10.1016/j.compbiomed.2025.110160
Raggio, C.B. et al. A privacy-preserving federated learning framework for generalizable CBCT to synthetic CT translation in head and neck. Front. Digit. Health 8, 1812254 (2026). https://doi.org/10.3389/fdgth.2026.1812254
Raggio, C.B., Bucher, L. et al. Security Analysis of a Federated Learning Framework for Medical Image-to-Image Translation. J Med Syst 50, 108 (2026). https://doi.org/10.1007/s10916-026-02436-8
Raggio, C.B. et al. Energy-efficient federated learning via adaptive encoder freezing for MRI-to-CT conversion: A Green AI-guided research. Sustain Comput Inform Syst 51, 101433 (2026). https://doi.org/10.1016/j.suscom.2026.101433
