Uncertainty-aware neural network within sequential federated learning for segmentation of diabetic macular edema

Abstract

The segmentation of spectral domain optical coherence tomography (SD-OCT) images for diabetic macular edema (DME) using deep learning technology has key challenges such as data privacy, computational cost, and information uncertainty. To address these, we present an uncertainty-informed neural network within sequential federated learning (UINN-SFL) for segmentation of DME. UINN-SFL organically combines sequential federated learning framework, feature discretization module based on rough fuzzy sets and adaptive genetic algorithm (AGA), and context pyramid fusion network (CPFNet) to reduce computational overhead and improve segmentation performance while ensuring data privacy. We compare UINN-SFL with mainstream SD-OCT fundus image segmentation algorithms on 100 3D retinal SD-OCT data with the gold standard. UINN-SFL outperforms other methods in all evaluation metrics. DSC, 95HD, and ASD of UINN-SFL are 0.8756, 0.8018, and 0.3118, respectively. The simulation experimental results demonstrate that our method can efficiently train models across different clients without data sharing, achieving accurate segmentation of DME.