基于条件生成对抗网络的视网膜黄斑分割  

Macula Segmentation Based on Conditional Generative Adversarial Nets

作　　者：傅迎华[1] 赵奇 FU Yinghua;ZHAO Qi(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《控制工程》2023年第6期1099-1104,共6页Control Engineering of China

基　　金：国家自然科学基金资助项目(61703277)。

摘　　要：黄斑在视网膜眼底图像中呈现为一个颜色暗淡的区域,且分割容易受到暗病灶干扰,因此提出一种可以快速准确分割黄斑位置的方法。首先,设计神经网络将深度学习的方法应用到黄斑分割中,并通过上采样过程恢复原始的尺寸;其次,受限于有限的数据集,将条件生成对抗网络(conditional generative adversarial nets,cGAN)引入到黄斑分割中,该方法可生成伪图像作为分割结果,并以一种可解释的方式对此结果进行训练来分割黄斑。实例验证结果表明,该方法可生成高质量图像,并能处理梯度消失的问题,有助于弥补数据样本不足的缺陷,在光照不均和暗病灶干扰的情况下,也能取得很好的分割效果。使用公共数据集MESSIDOR和Kaggle进行了训练和验证,分别得到了94%和99.32%的准确率。The macula appears as a dimly colored area in retinal fundus images,and segmentation is prone to interference from dark lesions.Therefore,a fast and accurate segmentation method for macula is proposed.Firstly,a neural network is designed to apply deep learning methods to macular segmentation,and the original size is restored through the upsampling process.Secondly,limited by a limited dataset,a conditional generative adversarial nets is introduced into macular segmentation.This method can generate pseudo images as segmentation results,which can be trained to segment the macula in an interpretable manner.The example validation results show that this method can generate high-quality images and handle the problem of gradient vanishing,which helps to compensate for the deficiency of insufficient data samples.It can also achieve good segmentation results which suffer from illumination variation and dark lesions.Training and validation were conducted using the public dataset MESSIDOR and Kaggle,achieving accuracy of 94% and 99.32%,respectively.

关 键 词：眼底图像 黄斑分割 条件生成对抗网络 糖尿病性黄斑水肿 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]