FedLiverNet: a federated learning framework for privacy-preserving and efficient liver cancer detection

Liver cancer continues to be a significant health issue on the global front, and proper segmentation of the liver and the tumor formed from the computed tomography is essential in the early diagnosis and subsequent treatment strategies. Although deep learning models can be trained to perform well in segmentation, the optimal way to train a strong model is with large, diverse datasets that may be distributed across institutions and cannot be centralized due to privacy and other regulatory restrictions. Federated learning enables joint training without exchanging patient data; however, performance may be poor on non-independent and identically distributed (non-IID) data, and privacy is a concern in optimization. This paper presents FedLiverNet, a communication-efficient and privacy-guaranteed federated liver and tumor segmentation system. FedLiverNet is a variant of the U-Net segmentation architecture that incorporates a modified backbone, differential privacy aggregation, and clustered federated learning with local adaptation to promote personalized support among heterogeneous clients. Simulation-based experiments indicate that FedLiverNet achieves a 0.89 ± 0.03 tumor Dice score and a 23% reduction in communication cost and is more effective than either federated averaging or local-only training under heterogeneous data distributions. These findings make FedLiverNet a viable solution to privacy-constrained, multi-center liver cancer detection and segmentation.