摘要
针对目前糖尿病视网膜病变图像人工分级困难、识别精度差等问题,研究了一种基于复合缩放模型的糖尿病视网膜病变图像分级方法,用于检测糖尿病视网膜病变,达到辅助医师诊断的目的。首先对糖尿病眼底病变图像进行了详细分析,经过图像预处理后,采用图像混合增强图像特征,同时使用图像翻转、图像加噪、调节对比度等不同方式扩充训练集,最后利用迁移学习和基于复合缩放模型方法进行糖尿病眼底图像病变程度分级。经过多组实验表明,该方法对糖尿病视网膜病变图像分级的准确率高达92%,Kappa系数高达0.88。所提出的方法无需指定病变的特征就能够达到高精度的病变分级,相较于AlexNet、LeNet、CompactNet、ResNet等模型两项指标有显著提升,对糖尿病视网膜病变诊断能提供有效的科学依据,同时对研究其他眼底病变图像也有一定参考价值。
In order to solve the problems of difficulty in manual classification and poor recognition accuracy of diabetic retinopathy(DR)image,a DR image classification method based on composite scaling model is studied,which is used to detect DR and assist doctors in diagnosis.Firstly,the image of diabetic fundus lesions is analyzed in detail.After image preprocessing,the image features are enhanced by image blending.At the same time,the training set is expanded by image flipping,image noise adding and contrast adjusting.Finally,the degree of diabetic fundus lesions is graded by using transfer learning and composite scaling model.After a number of experiments,the accuracy of the proposed method for diabetic retinopathy image classification is as high as 92%,and the Kappa coefficient is as high as 0.88.It is showed that the proposed method can achieve high-precision classification of diabetic retinopathy without specifying the characteristics of the disease.Compared with AlexNet,LeNet,CompactNet,ResNet and other models,the proposed method has significantly improved the accuracy of diabetic retinopathy image classification.The diagnosis of the disease can provide an effective scientific basis,and it also has a certain reference value for the study of other fundus lesions.
作者
蒋鹏
何勇
姚凯学
胡加德
JIANG Peng;HE Yong;YAO Kai-xue;HU Jia-de(School of Computer Science and Technology,Guizhou University,Guiyang 550025,China;Guizhou Kehai New Technology Development Co.,Ltd.,Guiyang 550002,China)
出处
《计算机技术与发展》
2021年第12期193-197,共5页
Computer Technology and Development
基金
贵州省科学技术科技成果应用及产业化计划项目(黔科合成果[2018]4002)。
关键词
糖尿病
视网膜病变
深度学习
迁移学习
特征增强
复合缩放模型
diabetes
retinopathy
deep learning
transfer learning
feature enhancement
compound scaling model