Research On The Classification Method Of Macular Edema In DR Fundus Images Based On Deep Learning

Diabetic macular edema(DME)is the main cause of vision impairment in diabetic patients.Therefore,the early detection of DME is helpful to reduce the risk of vision loss significantly.According to the DME international clinical classification standard,detecting and judging the presence of hard exudates(HE)close to the center of macula is a standard method to assess DME in the color fundus images.Therefore,macular center location and HE segmentation are the two core tasks of DME grading.Many methods have been proposed by researchers for the two tasks,but how to improve the accuracy of location and segmentation is still a difficult problem in fundus image processing.The traditional detection methods mainly rely on the accuracy of the extracted features.They have low detection efficiency and high time consuming.However,the deep learning offsets the shortcomings of traditional detection methods and has obtained breakthrough progress in medical imaging.Therefore,methods based on deep learning for macular center location and HE segmentation are proposed in this thesis.Finally,the classification of DME is achieved.Aiming at the task of macular center location,a method based on the improved Faster R-CNN and vessels marking to locate macular center is proposed in this thesis.Firstly,in the process of making the macular marking sample,the capillary characteristics in a specific area around the fovea are added,thus the characteristic of macula is enriched.Secondly,according to the size of marked macula,two different sizes of anchor are designed in the Faster R-CNN network to get a target reference frame that more suitable for the macula.It is shown that the modified Faster R-CNN can obtained more accurate central positioning of the macula by using the marked macula with capillary characteristics and the designed two anchors.Aiming at the task of HE segmentation,a method based on MD-ResNet is proposed in this thesis.This network combine dilated convolution and multi-level feature fusion strategy.Firstly,ResNet is used as the main structure.The residual block of the high-level network is changed into the dilated convolution residual block.Secondly,the detailed features of different scales in the low-level network are connected with the global features obtained by the high-level network through the skip connection.Therefore,a more refined effect of segmentation can be obtained.Based on the macular center location and HE segmentation,DME grading of fundus images is achieved based on the international clinical DME grading standards in this thesis.By establishing the fundus polar coordinate system and judging the position of HE in the fundus polar coordinate system,the fundus images are divided into four levels:normal,mild DME,moderate DME and severe DME.Finally,based on the previous work,the DME lesion intelligent analysis system is designed to achieve the purpose of automatic and rapid detection of fundus DME disease.The HEI-MED,e-ophtha EX and hospital datasets are tested by the proposed method.The results show that the test results have higher accuracy of location and segmentation,and finally a good DME grading is achieved.