Dual-Branch and Multi-Scale Information Complementary Network for Hard Exudates Segmentation in Color Fundus Images

Diabetic Retinopathy (DR) is known as one type of complication of long-term diabetes resulting in severe deterioration of vision, even blindness. Hard exudates (HE) are a type of diabetic macular edema and a prevalent cause of vision impairment for DR patients. Detecting HE is essential for DR diagnosis. Herein, a method for detecting HE with a dual-branch multi-scale approach is proposed. Specifically, the model consists of two branches: the main branch is designed to identify prominent features of the lesions, while the auxiliary branch utilizes a standard encoder and a ResNet50 encoder to identify deep features of the lesions, with two different resolution images serving as their inputs. Considering the variation in sizes and shapes, a multi-scale attention mechanism is integrated into the skip connections of the main branch, enabling the model to segment lesions of different scales. The feature maps finally obtained from both branches are aggregated before reaching the decoder, allowing the capture of contextual information from fundus images. Experiments on the IDRiD, E-Ophtha and DDR datasets show that the Dice of our method reached 74.53%, 68.87% and 59.40%, the IoU reached 59.40%, 52.52% and 42.25%, and the AUC reached 80.58%, 83.88% and 74.79% respectively.