Research On Retinal Image Multilesion Screening System Based On Deep Learning

In recent years,diabetes has gradually become one of the main diseases that damage human health.Diabetes has no obvious symptoms in the early stage so that it is difficult to be detected.However,diabetic patients are often accompanied by serious complications such as retinal exudation,hemorrhage and microaneurysm.So,the detection of these complications is helpful to early diagnosis of diabetes.Therefore,the research and design of an automatic screening system for diabetic retinopathy is of great significance.Exudation is one of the main types of diabetic retinopathy.The detection and analysis of exudation is of great significance for diabetic retinopathy screening and clinical diagnosis.Based on the improved u-net network,we add the local feature analysis based on image block and the multi output segmentation network designed with the concept of multi task learning.In order to better extract semantic information,we propose the improvement of concatenate path,residual block structure and multi-scale attention module to optimize the network structure.Finally,our improved algorithm achieves the results of AUC values of 0.95,0.97 and 0.92 on the HEI-MED dataset,DIARETDB1 dataset and MESSIDOR dataset,which is 1% higher than the original network on three datasets.By counting and calculating the area of exudates,the quantitative information of lesions was given.We also designed a multi lesion classification network based on the instance segmentation network Mask RCNN for exudation,hemorrhage and cotton-wool spots.The network can detect multiple lesions by extracting the prospect first and then fine judging.In order to make full use of the global information in the image,we introduce the Non-local module in the network to help extract the global features of the image.Experiments show that our network has achieved excellent experimental results in three kinds of lesions.Detection of exudative lesions has achieved AUC values of 0.961,0.984 and 0.943 in hei-med dataset,DIARETDB1 dataset and MESSIDOR dataset.Hemorrhage detection has achieved AUC values of0.927,0.954,0.941 and 0.963 in DIARETDB1,AI＿he1,AI＿he2 and AI＿he3datasets,respectively.While the accuracy of cotton-wool spots detection on DIARETDB1 is 0.8764.