Study On Classification Of Diabetic Retinopathy Based On Machine Learning

As one serious complicationof diabetes,patients who suffer from diabetic retinitis(DR lesions)are becoming younger and younger.Therefore,screening for the prevention and treatment of DR lesions becomes more and more important.A doctor's diagnosis mainly depends on his clinical experiences,the traditional method is based on the analysis of the main physiological symptoms with a lot of samples,then the classification criteria are set up according to the work done above.In the processing,these potential features are extracted from the fundus images data which collecting from the patients at first.In the end,classification methods in machine learning,such as SVM and Random Tree,are used to determine whether them belong to the right features or not.These means are time-consuming and labor-intensive,high cost.The early subtle symptoms of DR lesions are not easy to detect out,the treatment is not timely enough when these symptoms become obvious than before,the new method is urgently needed.Deep learning has a powerful ability of expression for non-linearity.Deep learning converts the original data to a higher level of expression through some simple non-linear model.Convolutional neural network(CNN),as one of the most important methods in deep learning,has a wide application in pattern classification.In this dissertation,I focus on the difficulties and problems mentioned above,and study the convolution neural network in the field of fundus image processing of DR lesions.The main contents are as follows:(1)Combined with the physiologicalcharacteristics of DR lesions reflected in fundus images,Ipropose a method for featuresextractingin fundus images.use edge detection,mathematical morphology and other methods to extract the blood vessel structure and detect out the exudation areas.Then the spatial registration is performed in the same image.At the same time,the difference between the original images is weakened,the main features are strengthened by usingimage fusion technology.(2)a convolution neural network model named DrNet is proposed for classifying the fundus images of DR lesions.Under the deep learning framework of CAFFE,DrNet is pre-trainedusing the ideasof Fine Tuning first.Finally,training and optimization are followed.The experimental results show that the model has a good effect on the classification of fundus images of DR lesions.