A multi-level context fusion network for exudate segmentation in retinal images

Abstract

Diabetic retinopathy(DR), is one of the major causes of visual loss. Exudates as the early symptoms of DR, are formed by leakage of fluid from injured retinal blood vessels. In clinic, exudate detection is mandatory for accurate analysis of progress of DR. However, because of the large intraclass variation and high interclass similarity, it is very challenging to detect exudate. To address this difficult problem, we propose a multilevel context fusion model that integrate the global abstract features and local refine spatial information simultaneously. Furthermore, the pyramid pooling module is introduced to gather the global information with varied scales. In addition, auxiliary supervision is exploited to alleviate the issue of gradient vanishing and enhance the representative capability of the lower layers of the network. The mechanism of transfer learning is employed to accelerate convergence and mitigate the problem of insufficient medical data. Moreover, compared with traditional approaches, there is no any preprocessing or postprocessing steps in our method, the network can be trained in an end-to-end way. Experiment results on the publicly available e-ophtha EX database illustrate that the proposed approach outperforms the state-of-the-art method with a fast speed, which makes it suitable for practical clinical applications.