一种融合残差模块的U-net对眼底图像渗出液的检测

Exudate Detection For Retinal Fundus Image Based On U-net Incorporating Residual Module

渗出液是糖尿病视网膜病变(Diabetic Retinopathy,DR)早期出现的特征之一,也是判断糖尿病视网膜病变严重性的重要指标.由于渗出液在亮度特征上与视盘具有很大的相似性,渗出液的分割通常需去除视盘的干扰.为了加强图像的细微特征,本文提出了一种新的图像增强方法预处理方法,使得渗出液比视盘具有更高的对比度.深度U-net在训练样本较少且对稀疏性目标分割时具有较好的效果.本文在U-net结构基础上,加入Resnet结构,通过增强细微特征的学习来改进渗出液检测.本文通过kaggle和DIARETDB1两个公开数据集来验证所提出的方法,其准确度、特异性和敏感性指标在两个数据集上分别为99.1%、99.3%、80.3%和99.0%、99.0%、89.3%.

Exudate is one of the early characteristics of Diabetic Retinopathy(DR),and also an important indicator for the severity of Diabetic Retinopathy.Due to the great similarity between exudate and optic disc in luminance characteristics,the interference of optic disc is usually removed in the segmentation of exudate.In order to enhance the fine features of the images a new image enhancement method was proposed to preprocess the images so that the exudate had a higher contrast than the optic disc.Deep U-VET has good effect on sparse target segmentation with few training samples.In this paper,based on the U-NET structure,the Resnet structure was added to improve the exudate detection by enhancing the learning of fine features.In this paper,kaggle and DIARETDB1 are used to verify the proposed method.The accuracy specificity and sensitivity indexes are 99.1% 99.3%,80.3% and 99.0%,99.0% and 89.3%,respectively.