Research On Segmentation Of Diabetic Retinopathy Based On Multiscale Deep Learning Model

Diabetic retinopathy is one of the ocular complications of diabetes.There are no obvious symptoms in the early stage.As the disease progresses,it will cause irreversible damage to the patient’s vision.Therefore,for patients with diabetic retinopathy,early intervention and treatment are necessary.In this paper,we focus on the automatic screening of diabetic retinopathy and the segmentation algorithm of lesions in fundus image lesions.The main work are summarized as follows:1.For the task of lesion segmentation in fundus images of patients with diabetic retinopathy.Considering the class imbalance between the lesions and the background,the small number of medical images annotated at the pixel level for segmentation tasks.We design a segmentation framework based on multi-resolution image patches and multi-scale feature fusion.Multiresolution image patches are obtained by setting different sizes sliding windows on the image.At the same time,a multi-scale feature fusion module with attention mechanism is designed to fuse the multi-scale features extracted from multi-resolution image patches.Morever,We propose a backbone based on Attention U-Net to extract image features and embed it into the above segmentation framework,establishing a lesion segmentation model MCnet for diabetic retinopathy fundus images.The backbone extracts multi-scale features of the image through a parallel atrous convolution module(IAC module),and connects IAC module with the feature map of the downsampling stage,thereby reducing the information loss caused by downsampling in the network encoding stage.Through experiment on the IDRiD dataset,the sensitivity and precision of MCnet for the prediction of microaneurysm,hemorrhage,hard exudate and soft exudate are better than some current fundus image lesion segmentation algorithms.Further through ablation experiments,it is verified that the segmentation framework based on multi-resolution image patches and multi-scale feature fusion proposed in this paper is suitable for various feature extraction backbone,and can achieve better prediction results.Simultaneously,it is verified that the multi-scale feature fusion method based on the attention mechanism is better than the method of taking the maximum value or the average value.2.Since the blood vessels in the optic disc of the fundus image and the hemorrhage,the optic disc and the exudate are similar in color,there will be misclassification when the lesions are segmented.Therefore,in this paper we propose an optic disc segmentation network(DCnet)to decrease its influence on lesion segmentation.The decoding structure of the U-Net network is improved,through the "mesh" connection,the features of each stage of upsampling are fused and sent to the attention module,so that the network can learn more valuable features while fusing multi-scale information.DCnet is trained and tested on three public datasets,REFUGE,RIM-ONE-r3,and Drishti-GS,it shows better segmentation results compared with existing models.Finally,combined MCNet with optic disc segmentation,an improved algorithm MCnet-OD is established for fundus image lesion segmentation.The images after DCnet optic disc segmentation are sent to MCnet for fundus image lesion segmentation,which improves the precision of MCnet for lesion segmentation results in the IDRiD dataset.3.Finally,we also establish a diabetic retinopathy screening and lesion segmentation system.First,the diabetic retinopathy screening network is established on the real dataset provided by the Endocrinology Department of Qilu Hospital by ResNet50.Secondly,the trained ResNet50,the optic disc segmentation network DCnet,and the fundus image lesion segmentation model MCnet are integrated into the diabetic retinopathy screening and lesion segmentation system.The input images are diagnosed with user interaction.The first step is to perform diabetic retinopathy screening.The input image is sent to the ResNet50 classification network after passing through the optic disc segmentation network to determine whether it has diabetic retinopathy.The second step is to segment lesions on the fundus images diagnosed as having diabetic retinopathy by the system,so as to draw conclusions for the reference of doctors.