Denissen, S., Laton, J., Grothe, M., Vaneckova, M., Uher, T., Kudrna, M., Horáková, D., Baijot, J., Penner, I.-K., Kirsch, M., Motýl, J., De Vos, M., Chén, O. Y., Van Schependom, J., Sima, D. M., & Nagels, G. (2026). Real-world federated learning for brain imaging scientists. Frontiers in digital health, 8, 1-13. https://doi.org/10.3389/fdgth.2026.1691088

@article{210db3dd28724fcdb774d0d9461eb9b4,
title = "Real-world federated learning for brain imaging scientists",
abstract = "Background: Federated learning (FL) has the potential to boost deep learning in neuroimaging but is rarely deployed in real-world scenarios, where its true potential lies. We propose FLightcase, a new FL toolbox tailored for brain research, and evaluate it on a real-world FL network to predict the cognitive status in patients with multiple sclerosis (MS) from brain magnetic resonance imaging (MRI). Methods: We first trained a DenseNet neural network to predict age from T1-weighted brain MRI on three open-source datasets: IXI (586 images), SALD (491 images), and CamCAN (653 images). These were distributed across the three centres in our FL network: Brussels (BE), Greifswald (DE), and Prague (CZ). We benchmarked this federated model with a centralised version. The best-performing brain age model was then fine-tuned to predict performance on the symbol digit modalities test (SDMT) of patients with MS (Brussels: 96 images, Greifswald: 756 images, Prague: 2,424 images). Shallow transfer learning (TL) was compared with deep transfer learning, in which weights were updated either in the last layer or across the entire network, respectively. Results: Federated training outperformed centralised training, predicting age with a mean absolute error (MAE) of 6.08 versus 7.02. Federated training yielded Pearson correlations (all p <.001) between true and predicted age of0.88 (IXI, Brussels), 0.91 (SALD, Greifswald), and 0.93 (CamCAN, Prague). Fine-tuning of the centralised model to SDMT was most successful with a deep TL paradigm (MAE = 9.19) compared to shallow TL (MAE = 11.05). Across Brussels, Greifswald, and Prague, deep TL predicted SDMT with MAEs of 10.71, 9.67, and 8.98, respectively, and yielded Pearson correlations between true and predicted SDMT of.25 (p = 0.282), 0.40 (p < 0.001), and 0.50 (p < 0.001). Conclusion: Real-world federated learning using FLightcase is feasible for neuroimaging research in MS, enabling access to large MS imaging databases without sharing data. The federated SDMT-decoding model is promising and could be improved in the future by adopting FL algorithms that address the non-IID data issue and consider other imaging modalities. We hope our detailed real-world experiments and open-source distribution of FLightcase will prompt researchers to move beyond simulated FL environments.",
author = "Stijn Denissen and Jorne Laton and Matthias Grothe and Manuela Vaneckova and Tom{\'a}{\v s} Uher and Mat{\v e}j Kudrna and Dana Hor{\'a}kov{\'a} and Johan Baijot and Iris-Katharina Penner and Michael Kirsch and Ji{\v r}{\'i} Mot{\'y}l and \{De Vos\}, Maarten and Ch{\'e}n, \{Oliver Y.\} and \{Van Schependom\}, Jeroen and Sima, \{Diana Maria\} and Guy Nagels",
note = "Publisher Copyright: 2026 Denissen, Laton, Grothe, Vaneckova, Uher, Kudrna, Hor{\'a}kov{\'a}, Baijot, Penner, Kirsch, Mot{\'y}l, De Vos, Ch{\'e}n, Van Schependom, Sima and Nagels.",
year = "2026",
month = mar,
day = "13",
doi = "10.3389/fdgth.2026.1691088",
language = "English",
volume = "8",
pages = "1--13",
journal = "Frontiers in digital health",
issn = "2673-253X",
publisher = "Frontiers Media",
}