Clustering Analysis Of Retinal Fundus Images For Diabetic Retinopathy

Diabetic retinopathy,short for DR,is a complication of diabetes.Neovascularization and preretinal hemorrhage occurring on the retina will lead to blindness in some severe cases.Based on the fundus images of diabetic retinopathy,some computer aided diagnosis systems constructed by deep learning have achieved competitive results and their decisions have got approval by professional doctors.However,such systems require a large amount of labeled data,which need ophthalmologists to mark,as training samples.Doctors’ labeling is mainly based on subjective experience which leads to a difference in doctors’ marking.And there are a variety of international classification standards for diabetic retinopathy.Different doctors use different classification standards to diagnose.Clustering analysis which can find data’s inherent rules through data driven was made to get preliminary annotations of fundus images.Clustering analysis can lighten the workload of doctors,and annotate unlabeled data objectively.Due to the great success of deep learning in feature representation,clustering analysis based on deep learning has become a hot research and application in recent years.This paper focuses on the traditional clustering method and clustering analysis based on deep learning to annotate unlabeled data.The main contents of the this paper include:(1)Traditional clustering analysis based on original images and image features were carried out.In the researches of clustering based on image features,two steps of feature extraction and clustering were mainly considered.For feature extraction,we developed a toolkit to extract common image features.For clustering,classical K-means and spectral clustering were compared.The K-means algorithm took into account four distance measures.Calculating the adjacency matrix is the key of spectral clustering method.This paper mainly considered two different construction methods of the adjacency matrix.(2)In the clustering analysis based on deep learning,this paper applied clustering methods based on deep transfer features and joint unsupervised learning of deep representations and image clusters.The clustering method based on deep transfer features used the features extracted by transfer learning for cluster analysis.This paper used the deep features extracted from the 17 th layer of VGGNet-19 to do clustering analysis.In the framework of joint unsupervised learning of deep representations and image clusters,deep learning used convolutional neural network,and clustering used hierarchical aggregation analysis.This study integrated two processes into a single model with a unified weighted triplet-loss function and optimized it end-to-end.In the cluster analysis experiments,the public data set Messidor and the data set ZSUDR collected by The First Affiliated Hospital,Sun Yat-sen University were used.There are four classes of Messidor datasets(levels 0-3),and ZSUDRs are of two classes.The above clustering methods were evaluated using two indicators of accuracy and purity.In the experimental results,the clustering method based on the deep transfer feature has been improved compared with the traditional clustering method.On the ZSUDR dataset,the accuracy and purity of the traditional clustering method were 50.24% and 0.6582 respectively,and the clustering method based on the deep transfer feature has an accuracy of 70.54% and purity of 0.8209.The experimental results demonstrate the effectiveness of the clustering method based on deep transfer features in the cluster analysis of fundus images of diabetic retinopathy.