EDU-Net: Retinal Pathological Fluid Segmentation in OCT Images with Multiscale Feature Fusion and Boundary Optimization

A team of researchers led by Zijun Lei has developed EDU-Net, a novel edge-guided dual-branch encoder-decoder network designed to improve the automatic segmentation of retinal pathological fluids in optical coherence tomography (OCT) images. The model addresses a critical challenge in managing diabetic macular edema (DME), the leading cause of severe visual impairment in diabetic patients. Current OCT-based segmentation methods struggle with the variable morphology of fluid accumulation and blurred boundaries due to noise. EDU-Net tackles this by integrating local details from EfficientNet's lightweight separable convolutions with global context from large-kernel efficient convolution (LKEC) modules, enabling precise detection of tiny lesions while capturing long-range dependencies.

EDU-Net's key innovation is its multi-category edge-guided attention module, which fuses high-frequency boundary detail information into each resolution feature to optimize segmentation boundaries. Extensive testing on both in-house and public datasets showed that EDU-Net achieves state-of-the-art Dice Similarity Coefficient (DSC) performance, particularly for intraretinal fluid (IRF) lesions, while maintaining efficiency and robustness. This advancement could significantly enhance clinical workflows for DME management by providing more accurate and automated quantification of retinal fluid, potentially improving patient outcomes through better monitoring and treatment planning.