The Reserach Of Deep Learning Model For Diabetic Retinopathy Grading

Diabetic retinopathy is a complication caused by diabetes mellitus.The pathological characteristics of diabetic retinopathy can be divided into grade 0-4 according to the international classification standard.Grade 0 is a completely healthy retina,and grade 1-4 can be divided into proliferative retinopathy and non-proliferative retinopathy according to the severity.The traditional grading method for diabetic retinopathy is manual grading using machine photography,including fluorescein angiography and optical coherence tomography.However,the high reliance on manual grading will lead to the waste of medical resources and a series of misdiagnosis and missed diagnosis.Therefore,with the rise of deep learning,some scholars have applied deep learning to the classification of diabetic retinopathy,but there is still a problem of low classification accuracy.In order to solve the above problems,the deep learning model for retinopathy grading was improved in this paper.The main contributions and innovations of this paper mainly include:(1)Firstly,the Residual Network(Res Net)was improved and the Squeeze-and-Excitation(SE)block was added in the front of the residual module to give a higher weight to the features defining lesions and enable it to extract features conducive to the judgment of lesions more accurately.Secondly,the form of extract and fusion characteristics of different scale features which belongs to Feature Pyramid Network(FPN)was used,so the network can extract more accurate to micro hemangioma,leakage,bleeding,and the size of the new blood vessels and so on the characteristics of the different scale.in the contest of kaggle diabetic retinopathy data set on the experiment in this paper,the classification accuracy of learning model reached 78.594%,and the kappa value reached 0.746.(2)For the retinopathy images with inaccurate classification results,the fusion weight mechanism and the improved Semantic Difference Maximization partial label classification algorithm(WSDIM)were used to classify them,when classification the set of labels which removing the wrong labels was regarded as the candidate label set.After experimental comparison on the data set,the classification accuracy of the model with partial labeling classification algorithm reached 80.039%,which was 1.8% higher than that without partial labeling classification,and the Kappa value reached 0.8,which was 0.054 higher than that before.In this paper,the improvement of classification model for diabetic retinopathy is improve classification accuracy and the kappa consistency,at the same time reduced reliance on manual labor,to a certain extent,reduce the misdiagnosis and misdiagnosis phenomenon.this model improve the retinopathy classification accuracy of grade 1,so it has a great significance in define whether the retina is pathological,which can help patients found early treatment,and on the great degree prevent diabetic retinopathy further deterioration.