Research On Optic Disc And Macula Location In Retinal Fundus Images

Lesions appearing in optic disc(OD)and macula area in retinal fundus,such as glaucoma,age-related macular degeneration(AMD)and diabetic macular edema(DME)will cause serious visual impairment.As important features in fundus image,optic disc and macula play important roles in the processing and analysis of fundus images such as disease location and segmentation.In the retinal fundus,the precise location of the optic disc and the macula is still very challenging due to uneven illumination,low contrast and pathological disturbances.In order to realize the location of optic disc and macula area rapidly and effectively,this thesis studied the location algorithm of optic disc and macula based on Prewitt operator-K-means algorithm and improved Faster R-CNN algorithm respectively,and compared and analyzed the experimental results of these two methods.The main work of this thesis is as followsIn the research based on traditional location method,a method based on improved Prewitt operator and K-means cluster algorithm is proposed to locate optic disc and macula.Contrast limited adaptive histogram equalization(CLAHE)and Gauss filtering are used to preprocess the original retinal fundus image first.Second,the image is scanned horizontally and vertically by using Prewitt gradient operator,and the region of interest(ROI)for location of optic disc is extracted according to morphological information.Third,the k-means clustering method is used to achieve the preliminary location of the macula Finally,the distance information between the optic disc and the macula is used to refine the location of the macula.In the research based on deep learning location method,we propose an improved Faster R-CNN based method is proposed for location both of optic disc and macula in retinal fundus Image.The proposed method consists of four modules:feature extraction network,region proposal network(RPN),context network,location and regression.Feature extraction network:In order to make the network more deeper and more capable of learning,residual network(ResNet)is used which adopts residual connection and batch normalization.Inception-v3 structure is adopted too.Region proposal network:Cascaded structure of RPN is proposed to refine the locations of RPN proposals and eliminate the influence of shallow details.Context network:A context network is proposed for the location because not only morphological feature but also relative location information should be considered.Location and regression:Smooth L1 regression loss function is used on the candidate box to achieve border regression.To evaluate the proposed methods,4140 images were randomly selected from the public datasets from Kaggle website,in which 3355 images were randomly selected as the training set for the deep learning network and the remaining 785 images were used as the test set both for deep learning algorithm and traditional algorithm.The location accuracy of the traditional algorithm for optic disc and macula on the test set were 96.87%and 84.32%,respectively.The accuracy of the deep learning algorithm for the optic disc and the macula were 99.49%and 98.09%,respectively.The experimental results show that the deep learning algorithm can locate the optic disc and macula effectively and has potential clinical application.