Research On Fundus Disease Related Object Detection Based On Color Retinal Image

Fundus image processing and analysis is the important application in medical field for the computer science.The screening of fundus is the effective approach for the fundus disease prevention and treatment though checking color retinal images.At the present stage,the screening of fundus diseases is conducted by ophthalmologist manually with analyzing the key objects(i.e.,retinal vessel,optic disc,fovea and lesions)in the fundus image.The manual screening is time-consuming,repetitive and tiring,and is impossible for the large scale purpose.Many patients hence become blind due to the late treatment.Therefore,automatically detecting fundus lesion related objects with image processing,computer vision and machine learning is important for computer aided diagnosis and the large scale screening of fundus diseases.This paper focus on retinal vessel segmentation,optic disc localization and segmentation,fovea detection,and fundus lesions detection and recognition.The main points are listed as follow:(1)To improve the performance of tiny vessel segmentation and overcome the interference of step edges in the fundus image,this paper proposed an iterative geodesic time transform(ItGTT)to segment retinal vessels.The vessel centerlines were first extracted by directional line detectors,and based on which non-centerline pixels were classified to vessels by computing geodesic time in an iterative manner.Vessel network were segmented successfully once the iteration is stopped.In addition,this work also proposed a phase congruency based method to accurately segment the low contrast or tiny vessels.The retinal vessels were first enhanced based on the phase congruency and then segmented based on the fully connected conditional random filed model.The experiments demonstrate the effectiveness of our proposed methods.The experiments demonstrate the effectiveness of our proposed methods.(2)To improve the robustness to fundus lesions and artifacts in the localization of optic disc(OD),this paper proposed a directional model,based on which the localization can be treated as a minimization problem.In this work,a global relaxed bi-parabola directional model(R-BPDM)and a disc directional model(DDM)were designed.For fusing the global and local cues,this paper then proposed a hybrid directional model to integrate the R-BPDM and DDM to localize the OD.In the segmentation of OD,to overcome the interference of lesions and bright pigmentations adjacent to OD and the low contrast OD boundary,this paper proposed a variational model with multiple energies to segment the OD.In the proposed model,a phase-based boundary energy,a PCA-based shape energy and a region energy were used to involve the active contour.The proposed method can segment OD boundary accurately.(3)For accurately detecting fovea in the fundus image,this paper proposes a multiply features fusion model to overcome the interference of lesions.In the proposed model,the first feature was extracted based on the global prior about the fovea,i.e.,the anatomical knowledge of retina,the second feature was defined based on the local appearance of the fovea in a vessel removed image,and the last feature was learned from the training images with a deep convolutional neural network for distinguishing the fovea and lesions in the fundus image.The proposed fusion model can accurately detect fovea in different databases.(4)Microaneurysms(MAs)are the first sign of DR.Since MAs are generally small,they are easily ignored by patients and ophthalmologists.To avoid the interference about vessels and artifacts,a candidate MAs extraction algorithm was devised by analyzing the gradient field of the image,and then the candidate MAs were segmented with fast level set method.Classifying the true MAs and the non-MAs is an extremely class imbalanced classification problem.Therefore,a class imbalance classifier was specially trained for the MAs classification.Furthermore,considering the dark and bright lesions are the saliency objects in the fundus image,this paper also proposed a visual attention model to detect these lesions in a unified framework.