TEDR-YNet: A Dual-Encoder Retinal Multi-Lesion Segmentation Network

Multi-lesion segmentation in diabetic retinopathy faces numerous challenges, including complex fundus structures, small inter-class differences between lesions, blurred lesion boundaries, and class imbalance. To address these challenges, this paper proposes TEDR-YNet, an innovative Y-shaped dual-encoder network for the segmentation of diabetic retinopathy lesions. This network utilizes a dual-encoder architecture to capture features from different representations, significantly enhancing the model’s expressive capability. In the bottleneck layer, a meticulously designed lesion-aware enhancement module is introduced to improve multi-scale lesion perception and boundary segmentation accuracy.Askip connection mechanism with gated convolution is innovatively incorporated to strengthen the model’s focus on lesion regions. An edge enhancement module is further employed to optimize boundary segmentation performance, and a hybrid loss function is designed to effectively mitigate the class imbalance problem. Extensive experiments conducted on three public datasets IDRID, DDR, and TJDR demonstrate the superior performance of TEDR-YNet, achieving average Dice coefficients of 68.61%, 58.51%, and 58.14%, respectively. These results represent improvements of 11.47%, 7.15%, and 9.8% over the baseline UNet model. Additionally, the mIOU metrics on the three datasets reached 53.24%, 46.84%, and 44.43%, respectively. These findings robustly demonstrate the precision and robustness of TEDR-YNet in multi-lesion segmentation for diabetic retinopathy, offering strong technical support for clinical diagnostic assistance.