Research And Implementation Of Automated Detection Of Exudates In Retinal Images

With the development of lifesyle and the improvement of living standards,diabetes has become a worldwide chronic disease,which impacts on people's health.Diabetic patients have hyperglycemia due to lack of insulin or cells' abnormal rejection to insulin.Persistent hyperglycemia disrupts the body's normal metabolic activity,causing metabolic disorders,and thus induce multiple complications.Among them,diabetic retinopathy(DR),which is one of the most serious complications of diabetes,has become one of the main reasons for adults' visual loss or even blindness.Therefore,the early diagnosis of DR and timely treatment are important for patients to reduce the risk of blindness.At present,the screening of DR is achieved primarily by ophthalmologists' manual examination,which is less efficient and relies heavily on the clinical experience of the ophthalmologist.Therefore,the study of fundus image's automatic detection to achieve the objective,efficient and accurate detection of fundus lesions has a positive significance.The presence of exudates is an early symptom of diabetic retinopathy,so early detection of DR can be achieved by detecting exudates.In this paper,the research on the automatic detection technology of exudates mainly includes two aspects:the detection method based on traditional computer vision and the detection method based on deep learning.The former can be divided into exudate rough segmentation stage and exudate fine classification stage.In the first stage,the morphological method is used to estimate the background,and exudate regions are obtained by morphological reconstruction and thresholding.In the second stage,the descriptors of regional structureness and the regional blobness are introduced,which are combined with other appropriate regional features to form feature vectors.The bagging decision tree based on ensemble classification is used to classify the regions in order to get the final result of exudates detection.The latter study focus on generative adversarial networks and propose an improved version of GAN to achieve a better detection of exudates in pixel-level.In this paper,the proposed automatic detection algorithm is tested on the public data set DIARETDB1.The sensitivity is 98.08%,the specificity is 94.39%,the positive predictive value is 94.07%,and the accuracy rate is 95.77%.After Comparing with other related work,we can find that the proposed method has certain advantages and potential application value.