Study On Fundus Image Enhancement Based On Controlled Diffusion And Adaptive Sparse Approximation

Human eye is an important visual sensory organ of human body and is the main source of information from the outside world.As the most important pho-tosensitive tissue in the optical system of the human eye,the retina,in particular the macular area,with fine structure and complex function,is vulnerable to ad-verse factors in vivo and in vitro,leading to the development of some fundus oculi diseases,such as age-related macular degeneration and diabetic retinopa-thy.As an important basis for diagnosing eye diseases and various diseases of the whole body,retinal retinoscopy plays an important role in ophthalmology.To obtain retinal images of the fundus through various ophthalmic medical devices,and to obtain the required qualitative or quantitative indicators,are the basis of ophthalmologic examination,diagnosis and treatment.Fundus retinal imaging is not only widely used in the diagnosis and treatment of retinal microvascular network abnormalities and macular diseases such as macular degeneration,and it can also be applied to the diagnosis of systemic diseases such as diabetes and hypertension.At present,the methods of obtaining fundus retinal images mainly include fundus photography,fundus fluorescence imaging and optical coherence tomography.It is of great significance for the early diagnosis of ocular diseases to visuliaze the normal structure and pathological changes in the macular region and quantify the thickness of nerve fibers.The emergence of OCT imaging technology makes it possible to obtain these important parameters rapidly and accurately.Recently OCT has been widely used in retinal in-vivo imaging,providing accurate basis for early diagnosis,clinical treatment and prognosis of various macular lesions.However,during the imaging process,fundus retina image is inevitably affected by various noises,blurs and data deviations,which reduces the resolution and image quality of the fundus image.Therefore,to denoise,enhance and segment the retinal image by advanced computer image processing technology for accurate measurement and identification of specific medical information,is a complex and important work to improve the level of fundus medical diagnosis and treatment achieving accurate medical needs,which is of great importance in guiding clinical work.The main work to improve the quality of ophthalmic images can be divided into two parts.One part is an improvement in imaging systems,structures and equipment,namely hardware processing.However,due to the particularity of the fundus imaging method and the system cost control,the improvement of the hardware alone can not completely improve the image quality and achieve satisfactory results.The other part is the subsequent processing of the digital images after imaging,i.e.software processing.In recent years,subsequent image processing,especially digital signal processing,has been greatly developed,signif-icantly improving the quality of images.Preprocessing images to improve image quality and readability has become an important direction of technical develop-ments.Therefore,the follow-up works of this paper mainly study and innovate the related algorithms of denoising and enhancement of ophthalmic images from the perspective of software processing.The researchers proposed many methods for denoising and enhancing the reti-nal images,such as gamma correction,histogram equalization,Laplacian sharp-ening,variational and partial differential equation methods,sparse and low rank approximation methods,and various filtering methods in transform domain,and the like.However,there are still many problems in denoising and enhancement of retinal images.The main challenges faced by these methods are,how to avoid noise enhancement,overshoot artifacts near the edges,excessive smoothing of image details and information loss.The variational methods and nonlinear partial differential equations,originally derived from physics and mechanics,have received extensive attention in the field of image processing with great success.This technique uses a variety of modern mathematical tools such as variational methods,partial differential equations,differential geometry,vector and tensor analysis,and computational fluid dy-namics.In the case of speckle noise interference,we use the controllable diffusion equation with flux correction transmission technology,and the curve evolution equation of the active contour with fuzzy control to enhance and segment retinal images.Local or non-local weighted averaging of images is an effective method for image denoising,where the choice of weights is critical.A.Buades et al.proposed a non-local mean(NonLocal Means,NL-means)filter based on the whole image information,which can effectively maintain the texture and detail of the image.When defining weights,this method not only considers the similarity of spatial positions between two pixels in an image,but also their similarity in gray level of the image.K.Dabov et al.proposed a more complex cooperative filtering method in transform domain.This method comprehensively utilizes similarity processing and multiple spectral transform methods(1D wavelet transform,2D wavelet transform and 2D DCT transform)to obtain better results.The sparse representation of the image comes from the "effective coding hy-pothesis." Mathematically,this belongs to the research of function approximation.Currently,researches on image sparse representations are generally carried out along two main lines.(1)Multiscale geometric analysis theory based on fixed basis functions.For example,various types of wavelet transform methods,and the like.This type of method performs a sparse representation of image by fitting its local geometry.(2)Sparse representation theory based on overcomplete dic-tionary.This kind of method obtains the learning dictionary from some samples through machine learning training,indicating that the basis function of the signal is adaptively selected according to the characteristics of the image itself,which can obtain much sparser image representation and has stronger adaptive ability.Low-rank approximation is a modeling method for recovering potential low rank structures from observed data.It is a tensor multi-directional sparse repre-sentation method.Low rank matrix approximation has made great progress in theory,algorithm and modeling,and has been widely used in computer vision and biological information processing with great success.In general,low rank matrix approximation can be divided into two types:low rank matrix decomposition method and rank minimization method.The low rank matrix decomposition method is a kind of widely used subspace learning technique,which decomposes the data matrix into the product of low rank matrixes.The rank minimization method uses regularization constraints to find the low rank structure of potential data.In many images,non-local self-similar blocks are redundant,and similar blocks have a high degree of correlation,which means that a data matrix com-posed of similar blocks has a low rank characteristic.Low rank property is a way of image sparse representation.Low rank matrix approximation provides an ef-fective method for modeling non-local self-similarity of natural images.Therefore,the filtering algorithm based on low rank approximation can effectively recover the low rank structure of the data from the noise data,and achieve the purpose of removing noise.In this paper,in order to adapt to the complex structure of retinal images,based on the sparse tensor approximation and the oriented diffusion of images,integrating geometric diffusion equations,low rank approximation,sparse repre-sentation,self-similar processing and feature adaptive processing techniques,we enhance,segment and measure fundus images extracted randomly from the Qilu Hospital of Shandong University,and the superiority of the proposed algorithm in retinal image enhancement was verified.Specifically,our main contributions include following four aspects:Firstly,an enhancement method of fundus photography based on controlled diffusion.For the fundus retinal image in ophthalmology with dim blur and noise interference,in order to improve the doctor' s ability to interpret tiny details of retinal images,we proposed a robust method for sharpening the fundus images based on a controllable diffusion equation.In our research,the changes of microangioma and soft exudation in the early stage of diabetic retinopathy were studied.First,the power law transformation was applied to the fundus image in an appropriate gray level range.Secondly,a self-similarity filter was used to denoise the transformed image,in particular,the region of interest needed to be extracted in the image.Finally,a controllable diffusion equation method was used to sharpen and enhance images,which highlights such important details as microangioma?hemorrhagic spots?exudates and so on,effectively avoids the artificial artifacts caused by excessive enhancement.The experimental results indicate that our method has important clinical significance in improving the clarity and resolution of fundus images and facilitating the accurate judgement of patient 's condition by clinicians.Secondly,a sparse joint filtering enhancement based on the classification of fundus data of optical coherence tomography.As research objects we random-ly selected OCT images of patients with elderly macular degeneration,diabetic macular edema,and macular laser burn.Through statistical analysis of optical coherence tomography data,we know that these data are subject to severe signal-related speckle noise,and the noise has the characteristics of non-stationary non-Gaussian distribution.We used the proposed classification method to segment optical coherence tomography data into four categories with different character-istics.Then,different image speckle filtering enhancement methods were used for different image feature regions.The proposed sparse joint filtering method was divided into two steps:initial estimation and final estimation,which was a cooperative filtering method from image block estimation to point estimation based on sparse prior,including image segmentation,cooperative filtering and aggregation.Compared with classic gaussian smoothing,median filtering and non-local self-similarity filtering,this method can effectively remove the speckle noise and maintain important details and layered structure features of retinal images,which is conducive to subsequent image measurement and quantitative analysis,and greatly improves the diagnostic levels of ophthalmologists.Thirdly,a low rank approximation filtering enhancement based on the classi-fication of fundus data of optical coherence tomography.As research objects we randomly selected OCT images of patients with cystoid macular edema,macular laser burn,and Vogt-Koyanagi-Harada syndrome.Similarly,based on statistical analysis and data classification for optical coherence tomography data,we pro-posed a low rank approximation filtering method.This was a cooperative filtering method from image block estimation to point estimation based on sparse low rank prior,which was divided into multiple iterative steps,each of which included im-age segmentation,singular value shrinkage,aggregation and back projection.The proposed low rank approximation algorithm effectively eliminates the influence of noise,and more effectively retain important information such as the retinal scar and fluid under retina and choroid atrophy,which provides a basis for subsequent image measurement and quantitative analysis,and is of great importance for the-evaluation of prognosis and treatment effect.Finally,an active contour segmentation method based on fuzzy control and its application in the classification and measurement of optical coherence tomog-raphy data.We used a fuzzy control strategy to design local energy terms suit-able for image gray inhomogeneity,and proposed an active contour segmentation method combining coarse and fine segmentations based on the fuzzy control,to achieve better robust and accurate segmentation,classification and measurement of different regions of optical coherence tomography data.Therefore,it is helpful for doctors of fundus disease to make accurate measurement and quantitative analysis of patient's fundus diseases,to guide doctors to make more accurate treatment,and to reduce patient's pain for a better service for patient.In a word,we used the proposed computerized image denoising,enhancement and segmentation methods to deal with the OCT images of cystoid macular ede-ma?age-related macular degeneration?diabetic macular edema?macular laser burn and Vogt-Koyanagi-Harada syndrome,which proves the effectiveness and advancement of these methods:they improves the contrast of the images,re-duces the noise in the images,enhances their main features,and divides effective pathological tissues for further quantitative analysis.As a domestic leading re-search with a number of independent intellectual property rights,this thesis will further strengthen the collaborative innovation among mathematics,information science and medicine,promoting the development of clinical ophthalmic medical technology and diagnosis and treatment level of doctors with ocular fundus.