Federated Learning for Brain Tumor Segmentation
Author Information
Author(s): Yahiaoui Mohammed Elbachir, Derdour Makhlouf, Abdulghafor Rawad, Turaev Sherzod, Gasmi Mohamed, Bennour Akram, Aborujilah Abdulaziz, Sarem Mohamed Al
Primary Institution: University of Echahid Cheikh Larbi Tebessi, Algeria
Hypothesis
Can federated learning with privacy-preserving techniques improve the segmentation of brain tumor lesions across multiple institutions?
Conclusion
The study demonstrates that federated learning can effectively segment brain tumor lesions while maintaining patient data confidentiality.
Supporting Evidence
- The federated model achieved DSCs of 89.85%, 87.55%, and 86.6% for whole tumor, tumor core, and enhancing tumor core, respectively.
- The study utilized the BraTS 2020 dataset, which is publicly available and contains MRI scans from glioma patients.
- Privacy-preserving techniques like differential privacy were implemented to protect patient data during model training.
- The federated learning approach allows for collaborative model training without the need for centralized data sharing.
- Results indicated that federated models maintained high accuracy with only marginal decreases in performance compared to centralized models.
Takeaway
This study shows that computers can help doctors find brain tumors better without sharing patients' private information.
Methodology
The study used a 3D U-Net model trained with federated learning on the BraTS 2020 dataset, incorporating privacy-preserving techniques like differential privacy.
Potential Biases
Variations in imaging protocols and patient demographics across institutions could introduce bias.
Limitations
The performance may be influenced by data heterogeneity between institutions and the study did not evaluate the practical implementation in clinical environments.
Participant Demographics
The study utilized data from multiple institutions, but specific demographic details were not provided.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website