ESTRO 2024 - Abstract Book

S4819

Physics - Quality assurance and auditing

ESTRO 2024

1522

Poster Discussion

Federated training of segmentation models for radiation therapy treatment planning

Christian Jamtheim Gustafsson 1,2 , Tommy Löfstedt 3 , Mattias Åkesson 4 , Viktor Rogowski 2 , Muhammad Usman Akbar 5,6 , Andreas Hellander 4,7 , Peter Larsson 8,9 , Annika Malmström 10 , Ida Blystad 11 , Anders Eklund 5,12,6 1 Lund university, Department of Translational Medicine, Medical Radiation Physics, Malmö, Sweden. 2 Skåne university hospital, Radiation physics, Department of Hematology, Oncology and Radiation Physics, Lund, Sweden. 3 Umeå University, Department of Computing Science, Umeå, Sweden. 4 Scaleout Systems, -, Stockholm, Sweden. 5 Linköping University, Department of Biomedical Engineering, Linköping, Sweden. 6 Linköping University, Center for medical image science and visualization, Linköping, Sweden. 7 Uppsala University, Deparment of Information Technology, Uppsala, Sweden. 8 Linköping University, Department of Health, Medicine and Caring Sciences, Division of Diagnostics and Specialist Medicine, Linköping, Sweden. 9 Region Östergötland, Center for Diagnostics, Medical radiation physics, Linköping, Sweden. 10 Linköping University, Department of Advanced Home Care in Linköping, and Division of Cellbiology, Department of Biomedical and Clinical Sciences, Linköping, Sweden. 11 Linköping university, Department of Radiology in Linköping and Department of Health, Medicine and Caring Sciences, Linköping, Sweden. 12 Linköping University, Department of Computer and Information Science, Linköping, Sweden Radiation therapy treatment planning can be time consuming, as it requires manual segmentation of the tumor and several organs at risk. In medical imaging, it has been demonstrated in many applications that deep learning segmentation models, such as the U-Net, can perform automatic segmentation in a few seconds. However, deep learning models with millions of parameters require large annotated datasets for supervised training, and the datasets need to be diverse in order for the models to generate good predictions for as many patients as possible. It is costly and time consuming to create large datasets, especially if the datasets should contain images and manual segmentations from several different hospitals. Many hospitals and researchers are reluctant to share their sensitive data, as it may violate GDPR or other legal or ethical regulations. The purpose of this work is therefore to demonstrate that it is possible to train segmentation models for radiation therapy treatment planning through federated learning, where a model can be trained without sending any image data between different hospitals. Instead, the segmentation model is locally trained at each hospital and sent to a global server, which combines all model updates and sends out a new model to all hospitals. This process is repeated until a good segmentation model is obtained. Federated learning has previously been used for brain tumor segmentation between some 60 sites. Compared to that work, our work is more challenging due to using full head volumes (instead of skull stripped volumes) and using manual segmentations from a clinical workflow (and the used guidelines for manual segmentations may vary between hospitals). Purpose/Objective:

Material/Methods:

Magnetic resonance (MR) volumes (T1-weighted, T1-weighted with gadolinium contrast, T2 FLAIR-weighted, T2 weighted with gadolinium contrast), a CT volume, and manual segmentations (gross tumor volume (GTV), clinical target volume (CTV), brainstem) were exported from the oncolocy departments of Skåne university hospital in Lund