基于迁移学习算法的糖网病自动筛选系统研究

Research on automatic classification system of diabetic retinopathy based on transfer learning algorithm

现阶段我国基层医疗条件落后于城市地区,具有基层的眼科医生资源匮乏和带标记的糖网病眼底图像数据不足等问题。针对以上问题本文提出基于迁移学习算法的糖网病自动筛选系统,受卷积神经网络中迁移学习思想的启发,通过构造出CNN+SVM的融合模型来实现正确的图像分类。本次试验采用源数据来自于山东省泰安某医院。相比于直接使用SVM模型,CNN+SVM融合模型训练和预测所需时间由1993秒下降到1518秒左右,kappa值由0.20上升到0.79,特异度和敏感度都提高了大概30%的幅度。结论:CNN+SVM的融合模型不仅加快了模型的学习效率,而且实现了高准确率的图像分类识别。

At present,the medical conditions of grassroots extremely lags behind urban areas in China. Lacking of resources for ophthalmologists with grassroots and the enough data of fundus images. In view of the above problems, this paper proposes an automatic screening system for diabetic retinopathy based on transfer learning algorithm, inspired by the idea of transfer learning in convolution neural networks, the correct image classification is achieved by constructing a fusion model of CNN ＋ SVM. The source data used in this experiment came from a hospital in Tai＇an, Shandong Province. Due to the use of the transfer learning algorithm,the time required for model training and prediction was reduced from about 1993 to about 1518 seconds, the kappa value rised from 0.20 to 0.79, and the specificity and sensitivity increased by about 30% compared with the directly use of the SVM model.The final results show that the CNN ＋ SVM fusion model not only accelerated the learning efficiency of the model, but also achieved high accuracy image recognition and classification.