Detection Of Changes And Segmentation Of Retinal Structures And Diabetic Retinopathy In Fundus Images

Computer-aided diagnosis(CAD)system for retinal fundus image analysis can screen all kinds of lesions,grade the situation of patients and provide the clinicians a lot of information for decisions.There are three critical problems discussed in the CAD system:image change detection captured at different phases from the same person,segmentation of anatomic structures,detection and segmentation of diabetic retinopathy(DR)lesions as well as diabetic maculopathy grading.It is a routine for the CAD need to compare a pair of images captured before and after the treatment during monitoring and observing the patients’ situation to help doctors to diagnose diseases exactly.Sometimes a long serial with many frames at different phases is captured and observed,and the CAD need to help doctors to detect the change regions and dynamic progress of retinal diseases by comparing the captured frames.It is necessary for clinicians that the CAD system can detect the lesions and grade the situation of the patient at the same time.Automatic diabetic maculopathy grading system requires to detect DR lesions as well as the anatomic structures such as optic disc and fovea.At the same time anatomic structures usually bring much distraction for DR lesions detection.Hence,image change detection on the image pair as well as the image serial with many frames,segmentation of DR lesions and anatomic structure are discussed and a diabetic maculopathy grading system is given in this thesis.The purpose of this thesis is to improve the performance of the CAD system and its practicability in clinic by investigating and analyzing these problems.The main points and contribution are made of the following section:1.For a pair of retinal fundus images,change detection methods based on sparse representation and low-rank decomposition is proposed.First use a background modeling method based on sparse local patches reconstruction to detect change of the pair of image.The background of the current frame is reconstructed by the sparse representation for some image-patch dictionaries.The proposed method is robust to the local intensity in images.The global intensity abruptness between the images is adjusted automatically by the sparse coefficients of the local dictionary,the distraction of local illumination variations on change detection results is removed.However this method is only good at detecting small change areas,it is difficult to deal with large change areas.Therefore,an improved method based on constructing low-rank image serial by linear interpolation is proposed and change detection is given by decomposing low-rank image matrix.The pair of images is expanded into a long serial after intensity correction and low-rank matrix decomposition is used to detect image changes.Linear interpolation makes image change detection of the image pair become low-rank matrix decomposition,reduces the illumination variation between the continuous frames and make change feature clear.The experiments in the second chapter shows the efficiency of the proposed method.2.For the long fundus image serial,a tensor low-rank decomposition model to detect the change areas is proposed.Although constructing an image matrix and leading to the low-rank Robust Principal Component Analysis(RPCA)decomposition keeps the succession in the serial,the contiguity of space is broken.Hence,a change detection method based on tensor RPCA is proposed in this dissertation to detect the change areas.This method keeps the succession as well as the contiguity at the same time.Total-variation is used to measure the temporal and spacial continuousness.Tensor RPCA is robust to abruptness of intensity and make the change mask clearer than matrix RPCA.Local tensor PG-TenRPCA is introduced in low-rank decomposition in this thesis,it is a local feature model,hence it is more robust to registration error than H-TenRPCA.3.For anatomic structures detection,a new method to detect and segment anatomic structure such as blood vessel,optic disc and fovea for retinal fundus image with lesions based on deep learning is proposed.Anatomic structure segmentation is usually distracted by DR lesions when there are lesions in fundus images.At first blood vessel detection based on deep learning is discussed and analyzed and then a new method based on deep learning with vessel constraint to detect and segment optic disc and fovea is proposed in this thesis.In order to remove the distraction of bright lesions,deep U-Net,Hough transformation and vessel constraint are combined together to detect the candidate location of optic disc.The proposed segmentation of optic disc is robust to the bright lesions’ distraction and can locate optic disc with good precision.For the low intensity of fovea and macula,the detection algorithms are usually sensitive to the local illumination.In this dissertion,a new method to locate fovea is proposed and the horizontal ridge is computed based on the vector sum of vessels at first.Then the location of fovea determined by the horizontal ridge and local neighbourhood searching.It is robust to illumination variation and can locate fovea with great accuracy at the end.4.For DR screening,a method to detect four kinds of DR lesions based on deep learning is proposed and an automatic diabetic maculopathy grading system is designed.A weight mapping is introduced to balance the samples to achieve more precision for lesion segmentation.A deep network based on U-Net is used to detect four different DR lesions is given and automatical diabetic maculopathy grading system is designed based on the aforementioned results.The grading system has got an acceptable performance in Messidor.