The Key Technology Research Of Fundus Image Processing Based On HSI Space

As many potential diseases of human body may be revealed from the retinal bloodvessels, fundus image processing is always the important and difficult in medical imageprocessing field. Clinicians can observe the experimental data on multi-level by usinggraphics technology, and this can help doctors analyzing lesions and other interested areaboth qualitative and quantitative. With the progress of science and medicine, color image isgradually becoming the mainstream of medical image processing,such as the ophthalmology.Therefore, it is proposed in this thesis how to improve retinal blood vessels in color fundusimages provided by hospitals as clear as possible to assist the clinical treatment. In this paper,we preprocess the actual color fundus images firstly, and then based on comparative studiesof various segmentation algorithms, one segmentation method can be exploited to achieve thefinal segmentation of blood vessels in fundus images. The main contents are as follows:Firstly, component selection in color fundus image processing. For color fundus images,we perform parametric statistics of each component for both RGB and HSI color space,before the pretreatment of multi-scale denoising is made respectively based on6componentimages. Through the edges-detected results of blood vessels used by the classic cannyoperator, the performance of6component images (R, G, B, H, S, I) for segmentation isproposed in the thesis. Both SNR of statistical parameter and experimental simulation resultsshowed that the I component in HSI color space can offer higher SNR and better robustness.It is said that the I component image is suitable for further retinal image processing.Secondly, choosing of the preprocessing algorithm. By the research on color fundusimages, it is confirmed that the difficulty of vascular segmentation is because of noiseplentifully, which can make the contrast of blood vessels and background deteriorate in thefundus image. The image may be bright in the middle but dark of all around. It is the matterfor incomplete vascular segmentation. In this paper, we combine some traditional noisereduction technology with multi-scale theory to choose the algorithm. Finally, the compoundalgorithm of vector median filtering and NSCT de-noising is adopted for image noisereduction, and the evaluation about the results is suggested by LMLSD algorithm.Experimental results indicate that the most obvious process result can be obtained by thedenoising algorithm because the SNR of the images is improved about8.50times compared with the original image by the estimation of LMLSD.Thirdly, the study on retinal blood vessels segmentation. About the image segmentation,two schemes, the traditional edge detection methods and the region-based segmentation, areimplemented. The results showed that although after a good denoising, the traditional edgedetection method is still sensitive to noise because of existence of false edges caused weakcontrast. Elicited by the previous research results about various component images in thispaper, we exploit iterative threshold method and OTSU method to segment blood vesselsfrom the fundus image. The experiment results show that a better result can be often achievedif OTSU threshold segmentation method is employed after we carry on high and low capalgorithm to enhance I component three times in HSI space. This can ensure the integrity andaccuracy of the fundus image blood vessel segmentation, and reach the purpose of auxiliaryclinical diagnosis.