Research On Segmentation And Classification Of Diabetic Fundus Image Based On Deep Learning

Diabetic retinopathy is a complication of diabetes with a high rate of blindness.It is mainly caused by retinal microvascular inflammation caused by hyperglycemia.Because its early symptoms are not obvious,it is often ignored by patients,and eventually leads to the loss of vision or even blindness.Therefore,regular diagnosis and screening of patients’ fundus conditions can greatly reduce the risk of blindness.During clinical diagnosis,doctors need to determine the degree of retinopathy of the patients based on the condition of fundus lesions and the characteristics of retinal vascular such as width,angle,and branching morphology.Then take corresponding treatment measures based on the degree of retinopathy.Currently,manual diagnosis is the main measure in clinical practice to determine the extent of lesions.However,due to the large number of patients with diabetes,and the characteristics of fundus images such as low contrast,complex retinal vascular morphology and structure,small fundus lesions and random locations,the manual diagnosis process is not only timeconsuming,but also may lead to misdiagnosis and missed diagnosis.Therefore,the automatic classification of diabetic retinopathy and the automatic segmentation of retinal blood vessels have important clinical application value.In recent years,with the continuous improvement of computer performance,deep learning has been widely used in various fields.Image recognition and understanding technology based on deep learning can independently learn the internal relationships of data,avoid subjective factors in diagnosis,and has the advantages of fast diagnosis speed and high accuracy.So it has become one of the research hotspots in the field of computer-aided diagnosis.Under this background,this paper have carried out research on retinal vascular segmentation and classification of diabetic retinopathy.The main work of this paper is as follows:(1)Aiming at the problems of insufficient sensitivity and weak generalization ability of existing models in blood vessel extraction,a dual decoder network model(ADNet)based on attention enhancement is proposed.This model inherits the codec ideas of U-Net.Firstly,the amount of training parameters is greatly reduced by reducing the number of convolutional filters in each layer of U-Net which can avoid model overfitting and improve the generalization ability of the model.Secondly,the sensitivity of the model is improved by adding an MFE module and an M/A intermediate decoder.As the first coding unit of the model,the MFE module can obtain rich vascular features under complex anatomical backgrounds.The M/A intermediate decoder consists of MFF and AHFF modules.The MFF module integrates deep semantic information and shallow spatial information to ensure full utilization of various scale features in the middle layer of the network.The AHFF module fuses hybrid features of different scales adaptively and generates two feature descriptors with different focus which can enhance the expression ability of the model.ADNet is evaluated on DRIVE,START and CHASE＿DB1 datasets.The AUC values are 98.41%,98.79%,and98.76%,respectively,and the sensitivity values are 84.19%,84.58%,and 82.62%,respectively.Compared to other end-to-end methods,ADNet has higher sensitivity and stronger generalization ability for vascular recognition.Compared with U-Net and its classical variants,ADNet has fewer parameters and higher segmentation accuracy.The model shows excellent comprehensive performance and has certain clinical application value.(2)Aiming at the problem that the classification accuracy of existing models is not high,a network model MFANet that focuses on local features and a network model ECARNet that enhances channel dependency are proposed,which are proposed according to the characteristics of lesions in diabetes fundus images,such as small size,diversify,and random of location.MFANet mainly solves the problem that existing models do not pay enough attention to small lesions.This model uses the improved Res Net34 as the backbone network,which reduces the amount of model parameters significantly by compressing the number of convolutional filters in each layer of Res Net34.Then,the local feature extraction network LFENet is used to improve the model’s ability to capture small lesions and enhance the proportion of detailed features in decision results.Driven by the dual networks of improved Res Net34 and LFENet,MFANet enables global and local features to participate in the diagnosis and grading of lesions.ECARNet mainly solves the problem of weak dependency relationships between feature channels in existing models.The model also uses the improved Res Net34 as the backbone network,but introduces an efficient channel attention residual module into the network.This module combines the residual module of Res Net34 with the efficient channel attention module ECA.It improves the interdependency between adjacent channels by setting local coverage,and avoids gradient disappearance during reverse propagation through residual connections.MFANet and ECARNet models are evaluated on APTOS2019 dataset.In the two-classification task of diagnosing whether there is diabetic retinopathy,the accuracy of the two models reaches 97.01% and 97.22%,respectively.In the five-classification task of diabetic retinopathy degree grading,the accuracy of the two models are 79.62% and 78.88%,respectively.Compared with classical networks and recent excellent algorithms,the two models not only have higher accuracy than other neural networks,but also have significantly lower training parameters than classical networks.