论文精读2-ReLayNet: Retinal Layer and Fluid Segmentation of Macular Optical Coherence Tomography·····

论文基本信息

标题：ReLayNet: Retinal Layer and Fluid Segmentation of Macular Optical Coherence Tomography using Fully Convolutional Networks
作者：Roy, Abhijit Guha，Conjeti, Sailesh，Karri, Sri Phani Krishna，Sheet, Debdoot，Katouzian, Amin，Wachinger, Christian，Navab, Nassir
来源：CVPR 2017
arXiv：https://arxiv.org/abs/1704.02161

摘要

Optical coherence tomography (OCT) is used for non-invasive diagnosis of diabetic macular edema assessing the retinal layers. In this paper, we propose a new fully convolutional deep architecture, termed ReLayNet, for end-to-end segmentation of retinal layers and fluid masses in eye OCT scans. ReLayNet uses a contracting path of convolutional blocks (encoders) to learn a hierarchy of contextual features, followed by an expansive path of convolutional blocks (decoders) for semantic segmentation. ReLayNet is trained to optimize a joint loss function comprising ofweighted logistic regression and Dice overlap loss. The framework is validated on a publicly available benchmark dataset with comparisons against five state-of-the-art segmentation methods including two deep learning based approaches to substantiate its effectiveness.

本文提出了一种新的深度学习网络结构ReLayNet，用于正常人和DME患者的OCT图像层次分割。ReLayNet借鉴U-Net的思想，分为下采样和上采样两个步骤，在训练过程中同时使用了交叉熵和Dice overlap loss函数进行优化。实验结果与5个state-of-the-art的分割方法进行比较，该方法具有更好的表现。

正文

本文要分割的层次：

最终结果是分为10个类别，其中7条边界+视网膜上面区域（RaR）+RPE下面区域（RbR）+Fluid。

网络结构：

Pooling与Unpooling：

损失函数：

Weighted multi-class logistic loss:

J=−∑x∈Ωω(X)gl(X)log(pl(X))

Dice loss:

Jdice=1−2∑x∈Ωpl(X)gl(X)∑x∈Ωp2l(X)+∑x∈Ωg2l(X)

权值的初始化：

ω(X)=1+ω1I(|∇l(X)|>0)+ω2I(l(X)=L)

其中 ω1 是为了减弱噪声对视网膜内部边界信息的干扰，提高分割准确率； ω2 是为了解决各个类别之间样本数量不平衡的问题。

网络训练

全局的损失函数：

Joverall=λ1Jlogloss+λ2Jdice+λ3∥W(∙)∥2F

其中 ∥∙∥ 为 Frobenius范数，为惩罚项目。训练的目标：

Θ∗=argminθ:{W(∙),b(∙)}Joverall(Θ)

通过其导数进行迭代更新。

整个训练过程如下：

实验数据

全部的数据来自杜克大学公布的10个DME患者个体的数据，共110张。其中每张图像都有视网膜8个层次的标定数据。点此链接下载。

实验分割结果

Fig. 6. Layer and fluid predictions of a Test OCT B-scan near fovea with DME manifestation, shown in (a) with the expert 1 annotations in (b), expert 2 annotations in (c), ReLayNet predictions in (d) and predictions of the defined 5 comparative methods in (e-i). CM-GDP and CM-LSE doesn’t include predictions for fluid. The fovea is indicated by the yellow arrow. The region with a small fluid mass is shown by a small white box

Fig. 7. Layer predictions of Test OCT B-scan with no fluid mass, shown in (a) with the expert 1 annotations in (b), expert 2 annotations in (c), ReLayNet predictions in (d) and predictions of the defined 5 comparative methods in (e-i).

实验分析

评价指标：
CE：estimated contour error for each layer分割的边界误差
Dice：Dice overlap score
MAD-LAT：the error in estimated thickness map厚度图的误差

与其它方法的对比：

本文方法不同参数的对比：验证skip connections、joint loss functions和depth of network:

交叉验证的结果：

黄斑中心不同区域的厚度差：