Research On Analysis Of Diabetic Retinopathy Images Based On Deep Learning

Diabetic Retinopathy(DR)is one of the leading causes of preventable blindness in adults.Performing retinal screening examinations regularly on all diabetic patients is the key to preventing blindness.However,the popularity of DR screening will bring about a rapid increase in diagnostic workload.The use of manual diagnosis by doctors is subjective and inefficient,and it can not meet the need of large-scale DR screening.Therefore,the retinal image computer-aided analysis technology is very urgent.Retinal image analysis is one of the hotspot in the field of medical image processing.Although retinal image analysis algorithm continues to improve,it has the problem such as complex feature extraction and poor generation.Standing from the actural requirement of a DR assisted screening system and aiming at helpfully improving screening efficiency,a thorough research in combination with deep learning algorithm on the methods of retinal image classification,vessel segmentation and red lesion detection is proposed in this thesis.Research results will provide technical support for DR assisted screening system.On the aspect of retinal image classification,due to the absence of labeled retinal image and the difficulty of deep neural network training,we propose a retinal image classification algorithm based on transfer learning and ensemble learning.Firstly,we use the pre-training model to initialize the parameters of the VGGNet16 and GoogLeNet models and fine tune them to reduce the training time.Secondly,we extract the feature vectors from the two models and train the SVM classifiers respectively.Finally,we use the average method in ensemble learning to average the classifiers to improve the robustness of the model.This method achieves high accuracy when testing on public datasets.Meanwhile,the DR lesion region is visualized by the weakly supervised learning algorithm.We also apply the weakly supervised learning algorithm to retinal image analysis,so we can visualize the DR lesions and locate the DR lesion by threshold segmentation.On the aspect of vessel segmentation,to overcome the shortage of vessel annotation of retinal images and overfitting in model training,we propose a vessel segmentation algorithm based on U-Net and ResNet.Firstly,we perform regularization,equalization and field of view extracting on retinal images to remove background noise.Then we use the sliding windows with overlapping area to generate image patches as the training data.Next,we add ResNet module to the U-Net model for preventing degradation of the model accuracy.We also combine the image segmentation evaluation index to optimize the loss function of model.Finally,we systhesize the image patches segmentating by the model.Results show that our method can improve the convergence speed and accuracy.On the aspect of red lesion detection,due to the small size of red lesions,it is easily distributed by blood vessels and noise.We propose a red lesion detection algorithm based on background estimation and CNN.Firstly,according to the label of the lesion,we use a overlapped sliding windows to intercept the image patch as training data and train the CNN model initially.Next,we use background estimation and veseel elimination to obtain lesion candidates.We then cut the positive and negative patch from the candidates according to the groundtruth.We also modify the loss function of the model to solve the class imbalance of dataset problem.The results show that our method can detect the red lesion in retinal image effectively.