Transfer Learning For Automatic Classification Of Diabetic Retinopathy With Fundus Images

Diabetic retinopathy(DR),a complication of diabetes,is a most common and serious diabetic eye disease that is preventable.In order to make a diagnosis,ophthalmologists mainly use the fundoscope to observe retina of the patient in clinical,which makes the diagnosis time-consuming,laborious and inefficient.To improve diagnosis efficiency,computer-aided diagnosis systems for DR have been developed.However,these systems were established using deep learning and a large number of fundus images.Since obtaining large amounts of data is very difficult and costly,it is a great challenge to develop a computer-aided diagnosis system with a few available data.Therefore,it is of great significance to develop a computer-aided diagnosis system for DR based on a small number of fundus images.Based on transfer learning,we carry out the research from two aspects using deep transfer learning features.(1)Comparing the classification performance for commonly used machine learning algorithms based on deep transfer learning features.When deep learning is used to solve the classification task with small datasets,it is easy to lead to overfitting and it is difficult to improve generalization ability.One of the solutions is to use neural networks,which is pretrained by a large number of data,to extract features from small datasets,and then train a traditional classification algorithm to complete the classification task.In this paper,we use a pretrained convolution neural network(VGGNet)to extract features from fundus images,and then compare the classification performance among eight classification algorithms based on these features.Experimental results show that support vector machine is the optimal classification algorithm.(2)Based on metric learning,we use siamese network to implement classification with a small number of training samples.When a very limited number of samples is available,the methods using pretrained neural networks to extract features that used for training traditional classification algorithms are not always effective.Inspired by one-shot learning,we first utilize deep transfer learning features extracted from large datasets to train the siamese network that we use,and then the network is used for classifying other small datasets.Experimental results show that the method proposed in this paper can effectively classify fundus images when the number of training samples is small.