Federated Generation and Fusion for Privacy-Preserving Cross-Domain Optic Disc and Cup Segmentation in IoT-Enabled Healthcare

Ocular diseases are a leading cause of permanent blindness worldwide, making early and accessible screening critical for vision preservation. Fundus imaging, increasingly deployed through Internet of Medical Things (IoMT) devices, offers a noninvasive and cost-effective modality for large-scale retinal screening in community and remote settings. AI-driven segmentation of the optic disc and cup from fundus images enables automated quantitative assessment for disease diagnosis. However, existing models suffer from severe performance degradation across heterogeneous imaging devices and raise substantial privacy concerns when centralizing multisite data for training. To address these intertwined challenges, we propose a federated generation and fusion framework for privacy-preserving cross-domain optic disc and cup segmentation in IoMT environments. The framework jointly addresses three complementary challenges: domain discrepancy is mitigated through generation-based image translation, privacy preservation is achieved via a Gaussian perturbation-based image fusion mechanism, and cross-site collaboration is enabled through federated optimization without raw data sharing. At each local site, a generative model bridges the domain gap via an ensemble of losses, including a fast Fourier transform (FFT)-based high-frequency loss for texture preservation, a multilayer contrastive loss for region-focused style transfer, a feature-matching loss for semantic alignment, and an adversarial loss, which are balanced through multitask learning. To protect local data privacy, a Gaussian perturbation-based, privacy-preserving image fusion strategy combines real and generated images using perturbed weights. The global segmentation model is then obtained via federated weighted averaging across distributed sites without raw data exchange. Extensive experiments on four public fundus datasets demonstrate that the proposed framework achieves competitive or superior segmentation performance across diverse target domains while providing empirical privacy protection. Results are reported as the average of three independent runs to ensure statistical reliability. The code is available at: https://github.com/JRHU0048/FundusGAN