Abstract:

Glaucoma is an irreversible blinding ophthalmic disease, which should be found and treated promptly. However, the process of artificial diagnosis is time-consuming and laborious. Furthermore, it is prone to missed diagnosis and misdiagnosis due to the limitation of basic medical resources. Therefore, auxiliary diagnosis of eye diseases using deep learning technology has become more significant. How to segment the retinal blood vessels more accurately and effectively become a hot research problem for auxiliary diagnosis of eye diseases. To this end, a novel network structure based on U-Net named Atrous Residual U-Net（AR-Unet） is proposed. In AR-Unet, the Residual Network（ResNet） is firstly introduced into U-Net for the sake of avoiding the vanishing gradient and the loss of image structural information. Then, the Atrous Convolution is integrated into U-Net for expanding the receptive field and improving the correlation between objects, which makes the segmentation of blood vessels more accurate. At last, extensive experiments are conducted on three public available color fundus images databases including DRIVE, STARE and CHASE. The experimental results demonstrate that the performance of the proposed AR-Unet method is better than that of well-known methods under different evaluation criteria.

Key words: retinal vessel segmentation, Atrous Residual U-Net（AR-Unet）, atrous convolution, U-Net, Residual Network（ResNet）

摘要：

青光眼是一种不可逆转的致盲性眼科疾病，应当早发现和早治疗。但人工诊断是费时费力的过程，而且受基层医疗资源的限制，人工诊断很容易产生漏诊和误诊的现象。因此，利用深度学习技术辅助诊断眼疾病具有重大意义。如何更为准确且有效地分割视网膜血管成为眼疾病辅助诊断的研究热点问题。于是，基于U型网络（U-Net）提出一种新的网络结构称为空洞残差U型网络（Atrous Residual U-Net，AR-Unet）。在AR-Unet中，为了避免U-Net中的梯度消失以及图像结构信息丢失等问题，将残差网络（ResNet）引入到U-Net中。为了扩大感受野和提高物体间的相关性，再将空洞卷积（Atrous Convolution）整合到U-Net中，从而使得血管分割更加精确。在三个公开的彩色眼底图像数据集DRIVE、STARE和CHASE上进行大量实验，结果表明在不同评价指标下，AR-Unet方法的性能均要优于大多数对比方法。

关键词: 视网膜血管分割, 空洞残差U型网络, 空洞卷积, U型网络, 残差网络