Classification Of Retinopathy Images Based On Deep Learning

The retina is a thin film on the inner wall of the eyeball with numerous microvascular networks.Under normal circumstances,the shape of the retina is always in a healthy state.However,some eye diseases such as diabetic retinopathy,cataracts and physical diseases such as hypertension,arteriosclerosis can cause some changes of the retinal structure(such as the width,angle,length of the blood vessels).Therefore,the analysis of retinal images is one of the main ways to diagnose various diseases.Due to the artificial features extraction and classification of retinal images is very dependent on the operator's existing experience,and the diagnosis of diseases through the naked eye may be misjudged.Therefore,the scientific analysis of retinal images is a key step in the correct diagnosis of various diseases.In recent years,thanks to the increasing maturity of deep learning technology and it has achieved breakthrough application results in the field of medical image processing.This paper will analyze and study the retinal images from classification and vessels segmentation based on the deep learning method.The main work of this paper is as follows:(1)The background and research status of retinal image classification and blood vessel segmentation technology are briefly introduced,and the working principle and use advantages of deep learning platform Caffe and TensorFlow are described in detail.The basic theory of deep learning and the key technologies used in this paper are introduced in detail:(2)This paper proposes an automatic classification system for diabetic retinal images based on deep learning technology.Firstly,the retinal images should be preprocessed before they are used as the training samples according to the characteristics of the retinal dataset.This preprocessing process mainly includes the retinal image denoising and data enhancement.Secondly,on the basis of AlexNet network,the batch normalization layer is introduced into the network to obtain a deeper convolutional neural network-BNnet.Then,this paper learn from the idea of transfer learning method,the BNnet network is pre-trained by using the ILSVRC2012 dataset and the obtained initial model is migrated to the enhanced diabetic retinopathy dataset for fine tuning in order to capture the depth features for classification.Finally,we designed a deep classifier-FCnet,which consisting of a fully connected layer.According to the features extracted by BNnet,the classifier can accurately classify the input retinal image into five categories:no diabetic retinopathy,mild diabetic retinopathy,moderate diabetic retinopathy,severe diabetic retinopathy and proliferative diabetic retinopathy.The experimental results show that the recognition rate of the proposed method on the test set can reach 93%,which is better than the existing retinal image classification method.(3)In addition,this paper also designed a retinal vessel automatic segmentation system.Firstly,according to the characteristics of the DRIVE data set,some pre-processing operations should be adopted for the retinal image,such as contrast limited adaptive histgram equalization,median filtering and normalization.And the retinal image and the corresponding expert labeled blood vessel image are cropped into sub-images of 48x48 size,the main purpose is to increase the number of training samples,reduce the size of the image and the scale of the neural network.Secondly,the local features and global features of retinal image are extracted by the low-level and high-level networks of the proposed full convolution neural network.Finally,the extracted features are interpolated and reconstructed by an upsampling operation to obtain a retinal blood vessel segmentation map.As the number of network layers increases,the retinal features extracted by the high-level convolutional layer will lose some details.Therefore,the simple upsampling operation will make the segmentation result tend to be rough.In order to make the neural network fully understand the characteristics of retinal images,this paper combines the features extracted from the lower convolution layer and the higher convolution layer of the network,so that the vessel tree segmented from the model is more realistic to the real effect.The specificity,accuracy,sensitivity,recall rate and AUC of the proposed method on DRIVE datasets are 0.9856,0.9512,0.7256,0.8768 and 0.9683,respectively,which are superior to the previous retinal vessel segmentation methods.It shows that the proposed method in this paper has great application prospects in computer-aided diagnosis of ophthalmic diseases.