Research On OCT Image Denoising Based On Sparse And Low-Rank Representations

Since the invention of optical coherence tomography(OCT)in the 1990s,OCT has gradually become a powerful imaging modality in the field of med-ical diagnosis.OCT is a noninvasive,nonionizing optical imaging modality,which has some advantages such as high security,high resolution and fast scanning speed.OCT is currently utilized in several medical and biomedi-cal applications including dermatology,density,oncology,and cardiology in addition to ophthalmology.High quality images are the basis for clinical diagnosis and image based postprocessing techniques such as feature extrac-tion and image segmentation.However,OCT images are inevitably suffer from speckle noise due to low coherence interferometry and the limitation of imaging equipment.Speckle noise severely degrades the quality of OCT images,consequently makes it particulary challenging to identify the fine fea-tures of object for clinical examination.Hence,suppressing speckle noise in OCT images is important to improve image quality and to further increase the accuracy of clinical diagnosis.However,how to remove speckle noise while preserving fine image features is a challenging task in the OCT image processing field.In the past few decades,great efforts have been taken for reducing speck-le noise in OCT images and a large number of approaches have been report-ed.Generally speaking,these methods can be divided into two categories:hardware-based methods and soft-ware based methods.However,hardware-based methods are not easily adapted to standard commercial OCT imaging systems because it requires significant modifications of the hardware of an existing imaging systems.An efficient and economical way of speckle noise reduction is software-based methods,which postprocess OCT images.These techniques focus on numerical algorithms in spatial domain or in a certain transform domain.Examples include Bayesian estimation method,diffusion based method,wavelet decomposition and robust principal component anal-ysis method.However,these algorithms do not take spatial redundancy of images into account or use a fixed basis to represent OCT images.Hence,it is difficult to extract signals with fine structures from the speckled OCT images.This thesis mainly focuses on the postprocessing method and the construction of the adaptive transform basis,and does several main tasks as follows:(1)This thesis proposed a speckle noise reduction method for OCT based on adaptive 2D dictionary.To reduce speckle noise while preserving local image features,taking full advantage of nonlocal self-similarity prior,nonlocal similar patches are first extracted from the noisy image and are put into groups using a gamma distribution based block matching method.Adaptive 2D dictionary is then learned for each nonlocal similar patch group.Unlike traditional vector based sparse coding method,we express each image patch by the linear combina-tion of a few matrices.This image-to-matrix method can exploit the local correlation between pixels.Since each image patch may belong to several groups,the depspeckled OCT image is finally obtained by aggregating all filtered image patches.Experimental results demonstrate the superior per-formance of the proposed method over other compared despeckling methods in terms of objective metrics and visual inspection.(2)This thesis proposed a speckle noise reduction method for multi-frame OCT data using multi-linear principal component analysis(MPCA).xviA natural image denoising method based principal component analysis is first introduced.In order to maintain image details and save running time,nonlocal similar 3D blocks extracted from noisy data are then grouped using k-means ++ clustering method.Transform basis matrices in three directions are learned for each 3D block group and MPCA transform is performed.Basing on the assumption that the energy of true signal will concentrate on a small subset corresponding to large transform coefficients,while the energy of noise will uniformly spread over the whole dataset corresponding to small coefficients,a shrinkage operator is defined by using linear minimum mean square-error estimation.Finally,the transform coefficients are shrunk to remove speckle noise and the filtered OCT volume is obtained by inverse MPCA transform and aggregation.Experimental results show the proposed method outperforms other compared approaches in terms of speckle noise reduction and fine detail preservation.(3)This thesis proposed a speckle noise reduction method for OCT based on tensor-singular value decomposition(T-SVD).Nonlocal self-similarity shows great potential in image denoising.The denoising performance can be obtained by accurately exploiting the nonlocal prior.Based on the study of low rank tensor decomposition,this thesis models nonlocal similar patches through multi-linear method and exploits the recently reported T-SVD.A defined tensor norm is used to characterize the information and structure of multi-linear data.Furthermore,we prove that T-SVD based nuclear norm minimization problem can be rewritten as low rank matrix approximation problem in the Fourier domain.Experimental results show that 3D image blocks can be represented by using T-SVD,and our method improves the efficacy of tensor denoising.(4)This thesis proposed two novel two-stage speckle noise reduction method for multi-frame OCT.Combining the advantages of the algorithms proposed in the previous chapters,two new two-stage denoising method is proposed.In order to main-tain the structural information of image,we seek adaptive tensor dictionary for the decomposition of each 3D block group in the first stage.For simplic-ity,we assume that nonlocal similar 3D blocks in each group share the same tensor dictionary,which has orthogonal property.Then,the adaptive ten-sor dictionary learning problem is modeled as an approximation between the observed noise tensor and the recovered tensor.Through some derivation,the solution of this problem is transformed into the 2D adaptive dictionary learning in the Fourier domain,which is mentioned in Chapter 3.The second stage of the proposed method I is the MPCA method proposed in Chapter 4.In the second stage of the proposed method II,higher-order singular value decomposition(HOSVD)is adapted to recovery the latent noise-free tensor,moreover,the Tucker rank is adaptive determined by Eckart-Young-Mirsky theorem.Compared with some related methods,the proposed methods have better denoising results.