A multi-scale gated network for retinal hemorrhage detection

Haiying Xia, Zengyan Rao, Zuoshan Zhou

Abstract: Retinal hemorrhage detection is of great significance for clinical diagnosis and disease control. However, most of the traditional methods need to obtain candidate lesions firstly, and then determine the true lesions. To address this problem, we propose an end-to-end multi-scale gated network (MGNet) to directly detect hemorrhage. Taken the U-Net as the backbone, we first add a skip connection gated module (SGate) to the skip connection to suppress useless information. Secondly, we propose a high-resolution and low-resolution multi-scale fusion module (HLMS) to improve the representation capacity of network through fusing information from the adjacent decoder layers. Furtherly, we propose a weighted Dice loss (W-Dice loss) to focus on the hard samples. Extensive experiments on two publicly available datasets: DIARETDB1 and IDRiD, demonstrate that the proposed MGNet achieves competitive results of hemorrhage detection compared with the state-of-the-art works.