Detection Of Diabetic Retinopathy Based On Deep Learning

With the improvement of modern people’s living standard and the change of lifestyle,more and more diseases begin to appear frequently and attract people’s attention.Among them,diabetes,as a disease that people pay more attention to,has an increasing incidence year by year.Diabetic Retinopathy(DR),as one of the major complications of diabetes,is one of the main means to detect diabetes.Therefore,how to detect lesions in the early stage and then carry out targeted treatment has become the main key task of doctors.Traditional DR detection is mainly composed of an ophthalmologist in fundus images of key focal area of the screen,and judgment,with the development of deep learning,DR detection based on deep learning more and more aroused people’s concern,but there is still a lack of training data,the network training time slightly long and cannot achieve clinical practical effect.Lesions in order to be able to get more detailed information,reduce the network training time,improve the scale of training data,this article proposes a double tower diabetic retinopathy detection model,through the two networks to complete retinopathy data and to training and learning after segmentation of data fusion of global and local lesion characteristics fully lesions characteristic information,and through the improvement of network loss function reference and attention mechanism,and through the improvement of network loss function reference and attention mechanism,improve the training precision of the network.The main work of this article is as follows:1.In order to improve the ability of image feature expression of the model,this article takes the Inception-V3 network with strong information acquisition as the main network architecture,migrates the model parameters to the DR detection task through the pre-training on the Image Net data set,further fine-tuning to adapt to the feature extraction of the lesions of the DR detection task,and reduces the network training time.At the same time,according to DR five classification difference between each kind of different characteristics,in order to increase the degree of differentiation between classes,the network loss function instead of the original cross entropy loss function and the mean square error of the joint damage,the article introduces the forecast in the cross entropy loss function and the real value of the distance change degree,improve the effect of distinguish between class.2.In order to obtain more comprehensive features of the image lesion area,an attention mechanism module combining space and channel was introduced.Compared with the ordinary attention mechanism module,this module not only considers the importance of pixels in different channels of the image,but also the importance of pixels in different positions under the same channel,thus improving the ability of the network to extract features of the lesion area.3.In order to better access to local details of quantitative characteristics of lesion,this article deals with data set segmentation,double tower classification detection model not only to the original image to extract the global characteristic,also on the segmentation image feature extraction,as a supplement to obtain local fine characteristics of lesions,solve the problem of lack of the original image in pixels in the process,thus increasing the expression of the characteristics of the test model.Finally,the improved model and method were tested on the data set of Diabetic Retinopathy Detection in the data modeling and data analysis competition platform(Kaggle)and the experimental results were analyzed.The results show that the method proposed in this article can effectively improve the accuracy of DR detection tasks and the class distinction.