Retinal blood vessel segmentation using fully convolutional network with transfer learning
Retinal blood vessel segmentation using fully convolutional network with transfer learning
复制标题
DOI：
10.1016/j.compmedimag.2018.04.005
复制
发表时间：
2018-09-01
影响因子：
5.7
通讯作者：
Ko, Seok-Bum
中科院分区：
文献类型：
--
作者：
Jiang, Zhexin;Zhang, Hao;Ko, Seok-Bum
由于视网膜血管被认为是眼科和心血管疾病诊断中不可或缺的元素，因此视网膜血管树的准确分割已成为自动化或计算机辅助诊断系统的前提步骤。本文提出了一种基于通过迁移学习预训练的全卷积网络的监督方法。该方法简化了典型的视网膜血管分割问题，从全尺寸图像分割到区域血管元素识别和结果合并。同时，额外的无监督图像后处理技术被应用于该方法，以细化最终结果。在DRIVE、STARE、CHASE_DB1和HRF数据库上进行了大量的实验，这四个数据库的跨数据库测试的准确性是最先进的，这也体现了所提出的方法的高鲁棒性。这一成功结果不仅为自动视网膜血管分割领域做出了贡献，而且支持了将深度学习技术应用于医学成像时迁移学习的有效性。
Since the retinal blood vessel has been acknowledged as an indispensable element in both ophthalmological and cardiovascular disease diagnosis, the accurate segmentation of the retinal vessel tree has become the prerequisite step for automated or computer-aided diagnosis systems. In this paper, a supervised method is presented based on a pre-trained fully convolutional network through transfer learning. This proposed method has simplified the typical retinal vessel segmentation problem from full-size image segmentation to regional vessel element recognition and result merging. Meanwhile, additional unsupervised image post-processing techniques are applied to this proposed method so as to refine the final result. Extensive experiments have been conducted on DRIVE, STARE, CHASE_DB1 and HRF databases, and the accuracy of the cross-database test on these four databases is state-of-the-art, which also presents the high robustness of the proposed approach. This successful result has not only contributed to the area of automated retinal blood vessel segmentation but also supports the effectiveness of transfer learning when applying deep learning technique to medical imaging.