Diabetic Retinopathy Detection

Diabetic retinopathy is a complication of diabetes.In long-term high glucose environment,retinal vessels will produce a series of pathological changes,such as micro aneurysms,hard exudate and soft exudate.According to the severity of lesions,DR can be divided into No DR,Mild DR,Moderate DR,Severe DR,Proliferative DR in five stages.About 50% patients with diabetes have some stage of the disease,resulting in visual impairment or blindness.How to distinguish between these stages accurately is a complex problem to be solved urgently.In this thesis,we take the algorithm competition of Kaggle as an opportunity.Based on the retinal fundus images and tags provided by the organizers,we study the application of blood vessel segmentation algorithm and deep learning algorithm.In this thesis,the main work is as follows:1.The representation of deep learning image features and the related theories of deep learning image classification are introduced in this thesis,including local features and deep learning.They give the theoretical basis for deep learning of diabetic retinopathy detection.The convolution neural network and the autoencoder are mainly introduced.2.In order to extract the fundus image features more accurately,remove the shooting angle and the interference of redundant pixels on the experimental results,this thesis propose an algorithm based on the Hough transform theory to extract the contour of the fundus image,so as to achieve the strict alignment of the fundus images.The experimental results show that the algorithm solves the problem of uneven eye image effectively and realizes the registration of dataset images.3.In order to express the lesions such as micro aneurysms,hard exudates and soft exudates better,this thesis propose a retinal blood vessel segmentation algorithm based on classification and regression tree and AdaBoost.Changes of the shape,diameter,size and branch angle of retinal vessels,and whether there are hemangiomas or exudates determine the period of disease directly.Therefore,it is very important of the blood vessels segmentation.The experimental results show that,compared to the unsegmented images,the classification accuracy and Kappa coefficient(consistency check index)of the segmented images have been improved.4.In order to improve the accuracy of diabetic retinopathy detection,this thesis propose a multi label classification model based on convolution neural network.We classify these fundus images by training an improved convolutional neural network and use the activation values of the last convolution layer to represent the features of fundus images.Then we remove the fully connected layer from pre-trained model and attach our own deep descriptor to the bottom.In the process of training,in order to compensate for the lack of training data,we use transfer learning method to train.By training and testing fundus image dataset provided by Kaggle platform,the Kappa coefficient is increased by five percentage point compared with the original VGG-Net model,and the accuracy of classification is increased by eight percentage points.