Background <p>Diabetic macular edema (DME) is a leading cause of visual impairment and blindness among the diabetic population, and leads to abnormal retinal morphology, distorted layer boundaries and blurred structures in optical coherence tomography (OCT) images. Accurate segmentation of retinal layers and pathological fluid regions is critical for clinical diagnosis, but remains challenging due to irregular fluid distribution and low boundary contrast. This study aims to develop an effective segmentation method to jointly extract retinal layers and fluid regions for assisting clinical screening.</p> Methods <p>A novel dual-decoder multi-task network with graph attention mechanism was proposed for joint segmentation. A primary decoder completed region segmentation, while an auxiliary decoder focused on boundary detection. A cross-decoder spatial attention module was designed for bidirectional feature interaction, and a global reasoning module was embedded to capture long-range anatomical dependencies. Experiments were conducted on the public Duke DME dataset with five-fold subject-independent cross-validation, and paired t-tests were adopted for statistical significance analysis.</p> Results <p>The proposed method outperformed comparative mainstream segmentation models in overall and category-wise evaluation. It achieved stable accuracy in normal retinal layer segmentation and obtained competitive performance in identifying fluid regions, effectively reducing the interference of pathological changes and improving boundary consistency of segmentation results.</p> Conclusions <p>The proposed method enables accurate joint segmentation of retinal layers and fluid regions. It provides a reliable automated analysis tool for diabetic macular edema, and can serve as an effective auxiliary reference for routine clinical screening and quantitative evaluation.</p>