Research On The Microaneurysm Detection And The Lesion Recognition In Diabetic Retinopathy

Diabetes is one of the most important chronic non-communicable diseases in the world.China is the hardest hit area of diabetes,and the number of diabetic patients in China has ranked first in the world.In recent years,with the increase in the prevalence of diabetes,the burden of disease caused by diabetes has become more and more serious.Diabetes and its complications have seriously affected the quality of human life.Diabetic retinopathy is one of the serious complications of diabetes,and proliferative retinopathy even has the risk of blindness.Microaneurysms are the earliest symptoms of diabetic retinopathy and play an important role in the initial screening of diabetic retinopathy.As the number of diabetic patients in China is increasing,and most of them live in underdeveloped medical areas,the diagnosis rate is low,and the awareness rate,treatment rate and control rate all need to be improved.Therefore,the use of computer image processing technology to realize the automatic detection of microaneurysms is of great significance for the early prevention and treatment of diabetic retinopathy.This article takes the published color fundus image as the research object,combines the anatomical structure of the eye and the pathological characteristics of the microaneurysm,and studies the microaneurysm detection and pathological identification of deep learning in diabetic retinopathy from two aspects: principle and realization method.The main research contents are as follows:1.In view of the high complexity of the current algorithms for microaneurysm detection,the large amount of model parameters and the long detection operation time,this paper proposes a new end-to-end fundus microaneurysm segmentation network.The network reduces the four down-sampling steps in the original U-Net structure to two,changes the deconvolution of the decoding part to up-sampling and convolution,and improves the layer-jumping splicing connection method to the layer-jumping minus connection method.The designed network structure has two advantages: first,the change of the layer-jumping connection mode is conducive to the network learning small target features that are not highly distinguishable and as much as possible to retain more suspected lesion areas;second,by reducing the convolutional layer in the original model greatly reduces the amount of model parameters,which is conducive to the lightening of the model and the shortening of the reasoning time of the model.The experimental results on the E-Ophtha-MA dataset show that the proposed model reduces the parameters of the original model to 5.41%,and reduces the detection time of microaneurysms in fundus images of1440×960 scale to 2 seconds.2.Aiming at the current poor detection effect of fundus microaneurysms,this paper proposes a weighted loss function based on Dice coefficient.In the process of network learning,using the loss function to reduce the loss weight of simple samples and increase the loss weight of difficult samples can guide the network to achieve more target area recall.In further experiments,it is found that reducing the maximum slope of the activation function of the last layer of the segmentation network helps to preserve the probability discrimination of the network output.Therefore,this paper proposes a long-tail activation function based on Sigmoid,which not only retains the normalization function of the original function,but also retains the probability discrimination of the network predicted image.A large number of experiments on ROC dataset and E-Ophtha-MA dataset show that the proposed method increases the average recognition rate of difficult areas by4.79%.It can be seen from the drawn FROC curve that the proposed method has strong generalization ability in microaneurysm detection,and has achieved more competitive performance.3.Although deep learning has the potential to quickly review many medical images and make diagnoses,its clinical interpretability is still challenging.This article uses the disclosed diabetic fundus lesion segmentation data set to form a healthy fundus image set and a fundus image set containing diabetic retinopathy in proportion.The Res Net-18 classification network is used to realize the two-class classification of the presence or absence of lesions.By visualizing the output of the convolution kernel during forward propagation and the weight coefficients in the network during back propagation,the pixel areas in the image that are sensitive to the classification results are highlighted.The classification and visualization experiments on the combined data set combining E-Ophtha-MA and IDRi D show that the classification network focuses more on the lesion areas that are more sensitive to the classification results during the learning process.