Research On Recognition Of Diabetic Fundus Lesion Based On Transfer Learning Hybrid CNN

Diabetic retinopathy is one of the complications caused by diabetes.The current clinical diagnosis methods rely on the experience of doctors,but in actual cases,clinical lesions exist near the blood vessels of the fundus with different shapes,and a slight error may lead to misdiagnosis.With the development of deep learning,more and more researchers are devoted to exploring methods to assist doctors in diagnosis.In clinical diagnosis,accurate segmentation of retinal vascular morphology is helpful for doctors to judge lesions.Therefore,it is important for the accurate segmentation of blood vessels.Taking the fundus retinal images from the DRIVE database as the research object,the images are enhanced by the weighted average grayscale processing method and the limited contrast adaptive histogram equalization method,and the images are amplified by flipping,rotating and translation methods,preserving the retinal image.The image has more effective information,which lays a good foundation for the subsequent segmentation work.The modified U-Net network is used to segment the blood vessels,and the segmentation result is linearized to enhance its readability.The results show that the average Dice coefficient,accuracy,sensitivity and specificity of this method reach 0.8247,98.72%,85.38%and 98.54%,respectively,which improves the segmentation accuracy.The use of retinal fundus images to detect the level of DR lesions can improve the efficiency of doctors’ diagnosis.Taking retinal images from the Kaggle Eye PACS database as the research object,the brightness of the image is enhanced by converting the image to HSI mode,and the image is resized and enlarged to prepare for subsequent network training.In order to improve the grading accuracy,a transfer learning method was adopted,the parameters obtained by the Inception-V3 network training on the Image Net dataset were retained,and then transferred to retinal images for training,so as to fine-tune the parameters and make them suitable for the task.Lesions were graded using SVM and Sotfmax.The results show that the accuracy rate of SVM classifier reaches 93.40%,and the accuracy rate of Sotfmax classifier reaches 92.04%,which improves the classification accuracy.Figure 43;Table 16;Reference 51...