Exu-Eye: Retinal Exudates Segmentation based on Multi-Scale Modules and Gated Skip Connection

This paper proposes an efficient method called Exu-Eye for hard exudate segmentation in retinal funds images based on multi-scale modules and gated skip connections. The key components of Exu-Eye are (1) two multi-scale modules; the first one is at the beginning of Exu-Eye, while the second one, Atrous Spatial Pyramid Pooling (ASPP), is inserted at the neck of Exu-Eye to improve the fundus image feature extraction; (2) ImageNet MobileNet encoder; (3) Gated skip connection mechanism to enhance the capture of more details of retinal eye exudate lesions. Different experiments have been conducted on publicly available datasets, namely IDRiD and Ophtha EX, to demonstrate the efficacy of our method. Exu-eye obtained 75.53, 83.54, 79.33, and 87.5% of recall, precision, F1, and AUPR metrics on IDRiD and 59.25, 61.59, 60.40, and 64.53% on Ophtha_Ex dataset, respectively. Exu-Eye also outperforms numerous state-of-the-art approaches.