Research On Analysis Algorithm Of Diabetic Retinal Images Based On Deep Learning

In recent years,the number of diabetic patients has gradually increased.Diabetic retinopathy that is one of the common complications of diabetes has also gradually increased.However,artificial diagnosis may be misdiagnosed or missed due to light fundus lesions,too many patients,and inexperienced doctors,which may prevent patients from receiving timely treatment and even cause blindness.Therefore,in this paper,a network model for the classification and a network model for lesion detection are designed for diabetic retina images to assist doctors in diagnosis.The main research contents are as follows:Aiming at the problems of image enhancement such as blurry,insignificant features,and weak contrast in datasets for training classification model and area detection model of diabetic retinal images,an adaptive enhancement algorithm of fundus images based on fractional calculus is designed,an adaptive enhancement algorithm of fundus images based on fractional calculus is designed.A new fractional order mask operator with different values in gradient directions and a variation of the center coefficient with differential order is designed.The fractional differential adaptive function is constructed by using the improved Otsu algorithm to determine the optimal gradient threshold T.Finally,compared with existing algorithm by the visual and parameters evaluation,the effectiveness of the algorithm in this paper is verified.Aiming at the problem of diabetic retinopathy classification,a new classification network model for classifying lesions of fundus images based on deep learning is designed.Based on the Inception-Resnet-v2 network model that combine the Res Net and Inception blocks,a new DSInception-Resnet model is designed.Then it is trained and tested using the Kaggle dataset.Finally,compared with the Res Net model and the Inception-Resnet-v2 model by weighted Kappa value.The results show that the designed DSInception-Resnet model can better classify the degree of lesions of fundus images.Aiming at the problem that key positions of specific lesion features such as micro-aneurysm,hard exudate,soft exudate and fundus hemorrhage cannot be detected after classification of diabetic retinopathy,a lesion segmentation model is designed to detect the lesion area of fundus images based on deep learning.Based on the U-net model,combining the advantages of Res Net that can increase the depth of the network while avoiding the gradual disappearance of the gradient,the Res U-net model is designed.Then it is trained and tested using the DIARETDB1 dataset.Finally,the parameters are compared with the U-net model and the existing algorithms.The results show that the designed Res U-net model can segment the lesion well and has universal applicability.