FUSION: Uncertainty-Guided Federated Semi-Supervised Learning for Medical Image Segmentation

Abstract

Federated learning (FL) for medical image segmentation poses critical challenges, including non-IID data distributions, limited access to labelled annotations, and stringent privacy constraints across institutions. To address these, we propose FUSION (Federated Unified Semi-Supervised Optimisation Network), a novel dual-path training framework that integrates both Federated Labelled Data Learning (FLDL) and Federated Unlabelled Data Training (FUDT). Central to FUSION is a two-stage pseudo-label refinement strategy designed to ensure robustness under real-world federated constraints. First, synthetic label denoising is performed using Monte Carlo dropout-based uncertainty estimation, enabling clients to identify and exclude low-confidence predictions. Second, prototype-based correction is applied to further refine pseudo-labels by aligning them with class-specific feature centroids, mitigating errors caused by domain shifts and inter-client variability. These refined labels are used for localised training on unlabelled clients, while a dynamic aggregation scheme modulated by a reliability-based hyperparameter μ adjusts the influence of labelled versus unlabelled clients during global model updates. This tightly coupled interaction between pseudo-label quality and federated optimisation ensures stability, accelerates convergence, and enhances generalisation across heterogeneous clients. FUSION is evaluated on three diverse datasets: TCGA-LGG (brain MRI), Kvasir-SEG (colonoscopy), and UDIAT (ultrasound) and consistently outperforms state-of-the-art FL models in Dice, IoU, HD95, and ASD metrics. Results confirm the critical role of synthetic label refinement in enhancing segmentation accuracy, boundary precision, and model scalability. FUSION provides a technically grounded, privacy-preserving, and label-efficient solution for real-world multi-institutional medical image segmentation tasks.