A Blockchain Enabled Federal Domain Generalization Based Architecture for Dependable Medical Image Segmentation

Federated Learning allows data to be trained together as a global model without leaving the local area. However, federated learning still suffers from central server failures and lacks a principle of trustworthiness. At the same time, in medical clinical practice, the federal learning model performs very poorly when generalized to healthcare institutions that do not participate in the common training of the model. Therefore, this paper proposes a trustworthy medical image segmentation architecture based on blockchain-enabled federated domain generalization. Based on federated learning, the reliability of the segmentation model can be improved based on enhanced model generalization by introducing a Fourier-based domain generalization mechanism and blockchain technology. We evaluated the proposed approach with the optic cup and optical disc segmentation task for retinal fundus images. The evaluation results show that the architecture outperforms fundamental federated learning in terms of model generalization and accuracy, while also providing reliability.