Abnormal Regions Detection In Retinal Images Based On Retinal Background Reconstruction

Retinal related lesions can damage the patients’ vision and may even lead to blindness.Early diagnosis can alleviate the progression of fundus diseases and prevent blindness.For the purpose,fundus disease screening is widely used by inspecting color fundus images of persons.Currently,many different techniques of computer-assisted lesion detection based on training samples or lesion features of a particular type of lesions are proposed.These methods can detect specific types of lesions from retinal images but often ignore other types of lesions.This paper studies the detection of various abnormal regions in retinal images,which is the key technology in computer-assisted reading retinal images.In general,the number and types of lesions in retinal images are unknown,and different types of lesions exhibit different visual characteristics.These factors make it difficult to detect various types of abnormal regions in retinal images.There are currently no effective ways to automatically detect various abnormal regions.Inspired by the doctors’ visual diagnostic mode,this paper puts forward an effective method to detect various abnormal regions.In this framework,we first collect a large amount of normal retinal images as prior knowledge,and then learn its background image from the test image under the guidance of prior knowledge,and thus separate various abnormal regions from the test image.The main contributions and innovations of this paper are summarized as follows:(1)A feasible computational method of detecting various abnormal regions from retinal images is proposed.By effectively suppressing the interference caused by different imaging condition bias and individual differences of retinal blood vessels,the detection problem of various abnormal regions in retinal images is transformed to the problem of retinal background image learning.Based on different characteristics of lesions and the background images in retinal images,two effective computational method for detecting abnormal regions based on background reconstruction is proposed: a method for reconstructing the background based on the bases or the dictionary,and another background learning method based on the low-rank decomposition.(2)A background learning method based on orthogonal base learning and coding is proposed.First,a set of orthogonal bases are learned from the normal set by Principal Component Analysis(PCA).Second,a test image is reconstructed based on orthogonal bases and the abnormal regions are detected by the reconstructed error between the test image and the encoded image.Finally,based on local visual context properties of lesions,the local visual perception model is used to further suppress the false positive regions.This method is computationally simple,and can effectively detect significant small lesions,but can not well detect large lesions.(3)According to the characteristics that the abnormal regions appear as sparse structures in the normal retinal image set,a hierarchical sparse background reconstruction method(HSL)is proposed.First,a dictionary is learned from a set of normal retinal images.Secondly,based on the sparse coding(SC)a test image,a plurality of background images with different approximation accuracies are obtained.Finally,the fine background image is further learned from these background images based on the low-rank approximation.This method uses dictionaries to sparsely encode test images and further learns fine background images from multiple complementary background images.HSL can learn better background images than the method in(2).However,HSL implicitly assumes that the lesions obey the Gaussian distribution when sparsely encoding,which is inconsistent with the actual distribution of the lesions.It can not distinguish between the lesions and the normal regions and therefore results in some false positives.(4)According to the statistical characteristics of lesions,a background reconstruction method based on dictionary learning and Mixture of Gaussian(DMOG)is proposed.This method uses a Mixture of Gaussian model to depict the distribution of the lesions.Meanwhile,the background image of the test image is reconstructed by MOG coding with the dictionary learned from the normal retinal image set.Since the MoG model can better describe the distribution of lesions,the normal regions are well approximated in the encoded background image,and the lesions are also well suppressed.Compared with the HSL method,it can better suppress the lesions.However,DMOG has made a tradeoff between fitting the lesion distribution and sparsely coding backgrounds,resulting in poor reconstruction of some personalized structures in the normal regions.Therefore,the proposed method can not well distinguish lesions from normal personalized regions.(5)Based on low-rank characteristics of the background image in the normal retinal image set,a two-step low-rank matrix factorization method for background learning is proposed.Different from the method of reconstructing the background image based on the bases or the dictionary,it first obtains the background images of different accuracies by the low rank approximation of the test images in different normal retinal image sets respectively.The test image is further subjected to a low rank approximation in the set of background images with different accuracies,and the test image is decomposed into sparse structures containing various types of lesions and a low-rank background image.Compared with the method of reconstructing the background based on the bases or the dictionary,the proposed method can better distinguish the weak lesions from the normal personalized regions.However,this method requires multiple low-rank matrix decompositions with high computational complexity.In addition,the method is highly heuristic and the model can be interpreted weakly.