Segmentation Of Hard Exudates From Retinopathy Based On Deep Learning

The eye is an important perceptual organ for the human body to observe the outside world.80% of the information received by the human body from the outside world is obtained through the eye.Diabetic retinopathy is not only one of the most common complications of diabetes,but also the most important manifestation of diabetic microangiopathy,which is the main cause of vision loss and even blindness.Fundus retina image is a kind of fundus image collected by color fundus imaging technology.Because this technique does no harm to the human body and can clearly observe the physiological structure of the fundus,it plays an important auxiliary role for doctors in the diagnosis of fundus diseases.By analyzing the fundus retinal images of patients with diabetic retina,doctors can determine the stage of diabetic retina,so as to determine the appropriate treatment plan.Among them,hard exudate,as one of the important standards in the early stage of diabetic retina,is of great significance for judging the staging of diabetes.Due to the complex physiological structure of fundus images,the diagnosis of hard exudates mainly depends on the manual labeling of ophthalmologists,which is not only time-consuming and laborious,but also easy to be disturbed by external interference.The traditional image segmentation methods are mainly based on the low-level features of the image,these algorithms lack stability and accuracy,and it is difficult to meet the requirements of clinical diagnosis.In recent years,the rapid development of deep learning technology provides a new method for fundus retinal image segmentation.According to the characteristics of hard exudates,a hard exudate segmentation algorithm from global to local is proposed in this paper.This algorithm can automatically learn the characteristics of hard exudates through iterative learning and solve the problem of low accuracy of traditional methods.At the same time,the learning method of feature fusion is adopted in the algorithm,which is helpful to enhance the stability of the model.The main work of this paper includes the following aspects:(1)the retinal fundus data set is analyzed,and the fundus data are preprocessed according to the characteristics of hard exudates.The fundus data set is processed by data preprocessing methods such as contrast enhancement,adjusting image saturation,hue,image rotation,and so on,and the data samples which are beneficial to network learning are obtained.(2)A hard exudate segmentation model based on global information features is proposed,and the segmentation of hard exudates is realized successfully.The model adopts the encoder-decoder structure,with the whole fundus image as input,the encoder structure is used to extract the feature information of the image,and the decoder recovers the feature information extracted by the encoder.In order to reduce the feature loss between encoders and decoders,we add a gap filling layer to the encoder and decoder structure,and the features encoded by the encoder go through the gap filling layer and then input the decoder structure.The segmentation results on the data set show that this method can effectively segment the hard exudate.(3)An image segmentation method based on patch is proposed.This method can effectively solve the loss of detail features of the image caused by the convolution kernel size of the convolution neural network.The patch-based segmentation method divides a complete fundus image into image sub-blocks of the same size,each image block and the image block retain theoverlap of pixels,and finally restore to the size of the original image through the stitching between the image blocks.In addition,in order to get accurate segmentation results,we add feature balance layer and non-local structure to the segmentation model.The balanced feature structure is used to balance the high-level and low-level feature information and reduce the loss of features.The Non-local structure strengthens the relationship between pixels and pixels in the image block.(4)A segmented network model with cascaded global and local features is proposed.In this model,the features of the image are extracted through a coding structure.In order to enhance the semantic features of the image,the structure of two sub-networks of global and local is adopted.The global network retains the context information of the image by learning the whole image.The local network uses the patch,corresponding to the whole image not only to solve the problem of small amount of fundus image data,but also to extract the detail features of the fundus image.Finally,the network combines global and local features,so it can better complete the segmentation task.In order to improve the accuracy of network segmentation,a multi-scale segmentation method is proposed.The multi-scale segmentation method mainly includes the following two aspects.One is the segmentation strategy from coarse to fine.First of all,the segmentation is carried out on the global network,and the gradual transition to the local network.Second,the segmentation methods of different scales are carried out on the two sub-networks.The hard exudate segmentation algorithms proposed in this paper are tested on the open data set IDRID and compared with other classical hard exudate segmentation algorithms.The results show that the algorithm has a good effect on the segmentation of hard exudates,which verifies the feasibility of the segmentation algorithm.