Research On Diabetic Retinopathy Classification And Lesion Detection Based On Deep Learning

Nowadays color fundus image processing and analysis is an important area in the field of medical image processing.The color fundus image is an important basis for the clinical diagnosis of eye-related disease by the ophthalmologist.By examining the potentially diseased areas in the fundus image thoroughly,an ophthalmologist can evaluate the patient's condition and give the appropriate treatment advice.However,due to lackness of medical resources and shortness of mannual diagnosis which is time-consuming and the effect is highly dependent on the doctor's experience,lots of patients can not receive timely treatment,which finally lead to irreparable vision impairment or even blindness.Therefore,automatically and accurately analysing color fundus images through methods of image processing,pattern recognition and machine learning and other fields plays a critical role in the early preventment and timely treatment.This paper mainly uses the method of deep learning to research the following three aspects:(1)Classification of the severity of diabetic retinopathy(DR).DR is the most frequently occurring complication of diabetes mellitus and a leading cause of human blindness.To automatically classify the severity of DR,this paper propose s a method which is based on deeply supervised nets(DSN)and deep residual networks(Res Net),and is named as “Deeply Supervised Res Net”.To combat the issue of class-imbalance in the dataset,we adopt cost sensitive learning and an oversampling method of cropping images.By introducing deeply supervised layers to intermediate hidden layers of a variation of Res Net,we can provide additional regularization during training the network.More importantly,we can perform multi-scale learning by leveraging the predictions of these intermediate supervised layers,thus improving the final performance of the network.(2)Detection of microaneurysms(MA).MAs are one kind of lessions caused by local distensions of the retinal capillary,and appear to be small red dots in fundus images.Because MAs are the earliest detectable DR-related lesion,its detection plays an important role in the clinical diagnosis of DR.Thus this paper proposes a method based on patch-level convolutional neural networks(CNNs)to detect MAs in fundus images.According to the priori knowledge of MAs,we firstly generate more than 40,000 fundus images with annotated MAs.After that,we crop image patches containing and not containing a MA from the generated images.Then we train a patch-level CNN using the cropped patches.When making a test on an unseen image sample,we use sliding-window method and get a pixel-wise prediction result.In order to further improve the detection performance of the network,we also use some simple post-processing steps to post-process the prediction results,since the predicted probability graphs contain many pixels with lower probability.(3)Weakly supervised lesion region detection.There is no other methods that have been proposed in this direction.Considering that in the actual scene of computer-aided diagnosis of DR,it is usually not enough to just predict the category of the sample,and is more important to predict accurately the location of lesion regions.In this paper,we try to investigate the research of weakly supervised localization of DR-related lesions.In detail,we first train a classification CNN of DR,then generate a heatmap of the target regions that are related to the category of the sample,and finally locate the lesion regions using the heatmap.