Research On Denoising Technology For Two-dimensional Gel Electrophoresis Images

Two-dimensional gel electrophoresis plays an important role in the development ofproteomics. Two-dimensional gel electrophoresis images are obtained by separating theproteins as spots on a gel, based on the isoelectric point and molecular weight of theproteins. Because of the huge amount of data, computer-based analysis of gel images is acrucial step in proteomic analysis. Computer-based analysis of gel images mainlyincludes image preprocessing (such as noise removal, background correction), detectionand quantitative of the protein, and gel matching, etc. And noise removal is the first andimportant stage of gel images analysis. Effective denoising can prevent theover-estimation of the image background and help extracting faint but significant spots. Itcan also prevent the artifacts (fake spots), thus resulting in more accurate determinationof spots. And it leads to more accurate estimation of spots properties, e.g. spot volume.This thesis mainly aims at exploiting the denoising technology for gel images.Spatial filtering of gel images is simple and has efficient computation. However, itintroduces blurring and severe distortions. Wavelet outperforms spatial filtering both inthe PSNR values and the image quality, but it produces the pseudo-Gibbs phenomenonaround the discrete points. In this thesis we use Translation Invariant (TI) denoising tosuppress the pseudo-Gibbs phenomenon by changing the location of the discontinuouspoints. Experiment results show that TI method outperforms wavelet, while it is limitedin the capture of the directional information. In gel images, protein spots have differentsizes and shapes, therefore we hope to obtain information from several differentdirections in order to avoid the loss of protein spots. Partial differential equations basedon Total Variation (TV) are anisotropic and have great advantage in obtaining thedirectional information and protecting the edges of gel images. However, the problem ofTV is the staircase effects, which can mix two adjacent protein spots on the gel image into one protein spot. This problem will directly affect the quality of the protein spotsdetection, quantification and matching. In this thesis a new denoising method, namedWTTV, is proposed by combining the above two methods effectively. WTTV has highability in edges and details protection. So it can reduce the false protein, avoid the loss ofprotein spots, and have a high PSNR value. Meanwhile, we proposed an improvedNon-Local means algorithm based on the similarity of the pixels. It has great advantagein protecting the structures, details and textures. And it also effectively solves theproblems of WTTV, such as sufficient parameters requirement, high computationalcomplexity, and practical application inconvenience, etc. The improved Non-Local meansalgorithm is simple, efficient, and feasible.The main innovations of this thesis are as follows:1. The WTTV method is proposed combining translation invariant denoising withpartial differential equations. It not only makes up the insufficient directional informationof wavelet, but also reduces the staircase effects. Experiment results with waveletdenoising, TI denoising, TV denoising and WTTV denoising are analyzed and comparedto show that WTTV outperforms other methods in both image quality and PSNR values.However, it needs sufficient parameters and has high computational complexity, which isinconvenience for practical application.2. The improved Non-Local means algorithm is proposed. This method is based onthe similarity of the pixels, and it has great advantage in protecting the structures, detailsand textures. The TI denoising, TV denoising, NL-means algorithm and the improvedNL-means algorithm are compared by experiment and the results show that the improvedNL-means algorithm can not only protect the weak protein spots, but also have highcomputational efficiency, which makes it suitable for practical application.