Cataract Opacity Detection And Nuclei Segmentation Based On Three-phase Level Set Method

With the development of science and technology,the amount of image data is getting larger and larger,and the information needed to be processed is more and more complicated.Therefore,the traditional manual processing of digital images cannot meet the demand.Automatic image processing technology has a wide application prospect in improving image processing precision,improving processing efficiency,and extracting complex image information.Based on Markov random field and level set method,this paper focuses on the biomedical application of image processing.New image segmentation algorithms are proposed for computer-aided diagnosis and extraction of key information in biological images.In the clinical diagnosis of cataract,the manual processing of retro-illumination images is time-consuming and subjective.When dealing with massive patients,it is urgent to propose an accurate and effective and automatic cataract opacity detection algorithm.This thesis presents two novel image segmentation algorithms based on Markov random field and level set method to detect cataract opacity in retro-illumination images.In addition,a nuclear segmentation algorithm based on the improved level set method is proposed to facilitate biological information processing.The main achievements and contribution of this thesis include:(1)A Markov Random Fields based algorithm aims to detect posterior subcapsular cataract opacity and cortical cataract opacity in illumination images is proposed.There are mainly two steps in the proposed method: cataract opacity detection and cataract type identification.In the opacity detection step,retro-illumination images are classified into image with severe cataract opacity and image with mild opacity or healthy image.Observation data of the image in two groups is extracted using different methods respectively.Markov Random Fields is applied to segment opacity in the image.In the cataract type identification step,watershed merging method is proposed to make the method robust to the influence of noise.Then mean gradient comparison model is proposed,and its result is combined with the result of Markov Random Fields to improve the accuracy of segmentation.The experimental results show that the accuracy of posterior subcapsular cataract grading in level 1～3 is 90.7%,83.8% and 91.3%respectively,which indicates that the accuracy of cataract classification has been improved.(2)An improved three-phase level set is proposed and it is applied to segment retro-illumination image.In Markov Random Fields segmentation method,the accuracy of cortical cataract opacity is not satisfactory because of the over segmentation caused by Markov Random Fields.Original three-phase level set method is improved in two aspects to decrease the influence of over segmentation: main gray scale is utilized instead of mean gray scale,and a gradient function is added in the energy function.Then the proposed method is applied in cataract opacity segmentation.It can segment retro-illumination image into background,opacity with high gray scale,and opacity with low gray scale.Compared with the original level set method and Markov Random Fields method,the segmentation result of the improved method contains less over-segmentation and the level set contour converges on the boundary of opacity edge better.The accuracy of cortical cataract grading is improved by the proposed method.(3)A novel algorithm to split overlapped nuclei is proposed based on the improved level set method.Level set method is proposed to extract contours and evaluate whether a nucleus is blurry or not.Candidate point is detected using different methods in these two kinds of nuclei.Then two algorithms are proposed to identify two types of concave point respectively.The contour segments are fuether assigned into different groups and segments in each group is fitted to an ellipse.The proposed algorithm achieves a high accuracy in splitting both blurred nuclei and clear nuclei.In nuclei image from curved substrate,which contains both blurred nucleus and clear nucleus,the proposed method can achieve an accuracy of 99.08%.Comparison study showed that the proposed method outperform other segmentation methods for splitting overlapped nuclei.