Study Of Diabetic Retinopathy Based On Deep Learning Method

Diabetic Retinopathy(DR)is a kind of retinopathy with specific changes,and it is also one of the most serious complications of diabetes.As the disease progresses,patients will suffer different degrees of vision loss.Early screening and treatment of diabetic retinopathy is the most important way to prevent patients from vision loss and blindness.Until now,the diagnosis of diabetic retinopathy mainly relies on the clinical experience of ophthalmologists.However,due to the population needed to be examined is so large,a wide range of diabetic retinopathy screening is difficult to carry out.Some patients will not be screened timely,which will lead to loss of vision and even blindness.Therefore,the automatic detection methods of retinal lesions in the fundus image has great importance to the large scale screening of diabetic retinopathy.This thesis propose a diabetic retinopathy screening method based on deep learning and ordinal classification.Firstly,the method use an effective pre-processing method to enhance the fundus image so that the lesion area in the image has higher visibility in the image.Then,for the problem of data imbalance,we mainly use the sampling method to solve it.Finally,the convolutional neural network combined with ordinal classification is designed to automatically grade the diabetic retinopathy.Experiments have shown that the proposed method is better than the method directly using convolutional neural network.In addition,although the above-mentioned screening method can achieve good performance,it still has some shortcomings.Because it is often difficult for the method to learn small differences in retinal images.Such as whether microaneurysms occured in the image.However,microaneurysms is the earliest visible symptom of diabetic retinopathy,and its detection is of great significance for the staging of diabetic retinopathy.Therefore,this thesis proposes a microaneurysms detection method based on convolutional neural networks with multi-scale lesion information.The method first uses the CLAHE algorithm to enhance the contrast between the microaneurysm and the background in the green channel of the retinal image.Then,for further shade correction,a series of filtering operations are used.Then,candidate regions of microaneurysm were obtained by thresholding methods.Based on candidate regions,microaneurysm candidate samples are then acquired in the enhanced RGB three-channel retinal images and the sample imbalance problem is handled.Finally,the detection of retinal microaneurysms was achieved through an independently designed convolutional neural network.The experimental results show that the detection method of microaneurysms proposed in the thesis are favorably comparable with the state-of-the-art.The method not only helps to grade the diabetic retinopathy,but also has clinical application value.