Automatic segmentation of retinal layer boundaries in OCT images using multiscale convolutional neural network and graph search
Automatic segmentation of retinal layer boundaries in OCT images using multiscale convolutional neural network and graph search
复制标题
使用多尺度卷积神经网络和图搜索自动分割 OCT 图像中的视网膜层边界
DOI：
10.1016/j.neucom.2019.07.079
复制
发表时间：
2019-11-06
期刊：
影响因子：
6
通讯作者：
Gao, Xieping
中科院分区：
文献类型：
--
作者：
Hu, Kai;Shen, Binwei;Gao, Xieping
光学相干断层扫描（OCT）图像中视网膜层的准确定量分析对眼部疾病的检测和诊断起着至关重要的作用。本文提出了一种将多尺度卷积神经网络（MCNN）与图搜索相结合的自动分割OCT图像中视网膜多层边界的方法。首先，我们提出了一种MCNN架构来提取视网膜层边界的多尺度特征，从而生成视网膜层边界的概率图。特别是，我们通过融合从不同大小的输入图像patch中提取的特征映射来构建MCNN架构，以学习视网膜层边界的多尺度信息。同时，我们根据位置信息区分背景像素，以降低网络将背景误分类为目标的概率。此外，我们提出了一种改进的图搜索算法来从概率图中检测最终层边界。最后，我们在公开的年龄相关性黄斑变性（AMD） OCT数据集上用八种最先进的方法评估了我们提出的方法。实验结果表明，该方法在定量结果和视觉效果方面优于其他最先进的方法。(C) 2019 Elsevier B.V.版权所有
Accurate quantitative analysis of the retinal layer in optical coherence tomography (OCT) images plays a crucial role in detecting and diagnosing ocular diseases. In this paper, we present a novel automatic method by combining multiscale convolutional neural network (MCNN) and graph search to accurately segment multiple retinal layer boundaries in OCT images. Firstly, we propose a MCNN architecture to extract multiscale features of retinal layer boundaries and thus to produce probability maps of the retinal layer boundaries. Especially, we construct a MCNN architecture by fusing feature maps extracted from different sizes of input image patches to learn multiscale information about the retinal layer boundaries. Meanwhile, we distinguish the background pixels based on the location information to reduce the probability that the network misclassifies the background as a target. Furthermore, we propose an improved graph search algorithm to detect the final layer boundaries from the probability maps. Finally, we evaluate our proposed method with eight state-of-the-art approaches on a publicly OCT dataset with age-related macular degeneration (AMD). The experimental results demonstrate that the proposed method outperforms other state-of-the-art approaches in terms of quantitative results and visual effects. (C) 2019 Elsevier B.V. All rights reserved.