Retinal vessel segmentation via multiscaled deep-guidance

Abstract

Retinal vessel segmentation is a fundamental and crucial step to develop a computer-aided diagnosis (CAD) system for retinal images. Retinal vessels appear as multiscaled tubular structures that are variant in size, length, and intensity. Due to these vascular properties, it is difficult for prior works to extract tiny vessels, especially when ophthalmic diseases exist. In this paper, we propose a multiscaled deeply-guided neural network, which can fully exploit the underlying multiscaled property of retinal vessels to address this problem. Our network is based on an encoder-decoder architecture which performs deep supervision to guide the training of features in layers of different scales, meanwhile it fuses feature maps in consecutive scaled layer via skip-connections. Besides, a residual-based boundary refinement module is adopted to refine vessel boundaries. We evaluate our method on two public databases for retinal vessel segmentation. Experimental results show that our method can achieve better performance than the other five methods, including three state-of-the-art deep-learning based methods.