| Diabetic retinopathy is a complication caused by diabetes,which is mainly manifested in ocular angiogenesis.Because there is no significant change in the early stage of diabetic retinopathy,and there is little difference between the grades of diabetic retinopathy.The condition develops rapidly,which will cause the condition to worsen for a long time,eventually leading to vision loss and blindness.At present,clinical ophthalmologists make diagnosis and treatment after finding the location of the lesion based on the retinal image.This depends on the experience of the clinical ophthalmologist,which often leads to missed diagnosis and misdiagnosis.Coupled with the limited medical conditions in the area and the low efficiency of artificial retina image processing,patients have delayed the optimal treatment time.Therefore,under the influence of computer aided technology.This paper combines image processing technology and deep learning methods to analyze retinal images and build a series of different task models of diabetic retinopathy images,so as to achieve the role of assisting clinical treatment.According to the characteristics of retinal images,the analysis of retinal images is combined with deep learning methods,including the classification of diabetic retinopathy,the detection of lesions and the segmentation of retinal blood vessels.The main research work is as follows:(1)This paper proposes an automatic classification algorithm for diabetic retinopathy based on the A-ENet model.The algorithm uses the EfficientNet network to adaptively optimize the three dimensions of the width,depth and resolution of the network to improve classification performance.It introduces the attention mechanism in the process of network feature extraction to obtain the weight information of lesion features and weaken the non-lesion areas information to learn about the features of the lesion.The A-ENet classification model solves the problem that it is difficult to subdivide because of the small difference between the lesion categories.This model has played a positive role in theclassification of diabetic retinopathy.The classification accuracy rate on Kaggle and other data sets reached 97.2%,and the quadratic weighted kappa value reached 84.0%.(2)For the detection task of microaneurysms,the microaneurysms are unequal in size and unevenly distributed in the image.This paper improves on the basis of the regionally-nominated target detection model Faster R-CNN to generate anchors,and proposes a microaneurysm detection model based on the adaptive generation anchor mechanism.By introducing an adaptive anchor mechanism based on the feature map of RPN,the problem of too low number of anchors leading to low detection recall and the problem of feature map matching the shape of the anchor are solved.The detection model generates a candidate frame suitable for the size of the microaneurysm,so that the microaneurysm can be accurately detected.The detection model was tested on the public IDRiD data set,and the detection accuracy reached 90.2%,which verified the efficiency of the model.(3)Aiming at the problem of thin vascular structure and high curvature in the retinal image vascular segmentation task,which resulted in low segmentation accuracy,an R2U-Net segmentation model based on pixel attention mechanism was proposed.By adding the attention module on the basis of extracting deep features from R2U-Net,the attention mechanism is used on the jump connection,so that the down-sampled feature map of the same layer and the up-sampled feature map of the previous layer are weighted by pixels.Attention to blood vessel information is achieved,and accurate segmentation of blood vessels is achieved.Through experiments on the DRIVE,STARE and CHASE data sets,the average accuracy,sensitivity and AUC reached 96.5%,79.8% and 98.2%.Based on the pre-processing of retinal images,the deep learning method is used to analyze and study the diabetic retinal images.The experimental results verify the effectiveness of the proposed method. |
PDF Full Text Request