Study And Application Of Segmentation Methods For Fundus Images

Fundus image segmentation technology has always been a hot and difficult problem in medical image processing. Although a lot of research work has been done on it, there are still many new challenges expected to be solved with the urgent demand of clinic detection and the disadvantages of existing methods. Study on the fundus image segmentation methods, which can satisfy the clinical examination requirement of accuracy, objectivity and repeatability, has great signification for clinical research, diagnosis and treatment. Under this background, the dissertation has deeply studied the fundus image segmentation methods in theory and practice, and the main contributions and innovation are as follows:(1) Image denoising and field of view (FOV) extraction in fundus image preprocessing are explored, an image smoothing method based on adaptive median filtering and a fundus FOV extraction method based on HSV space model are realized. Adaptive median filtering changes the filter window with the variation of noise density, and takes different solutions to deal with noise and pixels. Therefore, this method can better preserve the edge and detail pixels as noise removing. Fundus FOV extraction method transforms the RGB image into HSV space model, then plots the histogram of value image and extracts the whole fundus FOV. These methods provide support to the subsequent processing.(2) The operation mechanism and dynamic behaviors characteristics of improved pulse coupled neural network (PCNN) are analyzed, and according to the dynamic behaviors, the rule of the improved PCNN parameters determination is proposed. Furthermore, by introducing the improved PCNN into fundus image segmentation, two segmentation methods based on human vision system are proposed. One is new segmentation method combined Otsu algorithm with improved PCNN, and the other is new segmentation method based on distributed genetic algorithm (DGA) and improved PCNN. Experimental results show that these methods applied to fundus image segmentation are feasible and effective.(3) In view of the disadvantage that global single threshold methods can hardly segment objects from the gray value overlapping domain of objects and background, a new segmentation method based on transition region extraction is proposed. The main vessels are segmented through optimal entropy method and the transition region is extracted by algorithm based on distributed genetic algorithm and dual Otsu. The final results are obtained via analyzing the region connectivity. Owing to the utilization of gray value distribution and regional configuration characteristics, experimental results indicate that the proposed method outperforms the Hoover algorithm on the small vessels extraction, connectivity and effectiveness. In addition, the efficiency of proposed method could be improved by introducing the distributed genetic algorithm based on migration strategy.(4) Considering the challenges of segmenting blood vessels in pathological fundus images, a new method based on divergence of vector field and directional local contrast is proposed. Firstly, the divergence of the vector field is used to locate most centerlines of pathological fundus image. Then the directional information of each pixel in centerlines is computed and the pixels around the centerlines are detected by modified directional local contrast method. Finally, the whole blood vessel network is obtained via reverse tracing at the end of each blood vessel segment. Experimental results show that the proposed method is robust for all kinds of pathological fundus images and has clinical reference and practice value.(5) A new method for measurement retinal blood vessel widths based on prior knowledge is proposed. First, the blood vessel segmentation result is used to extract the skeleton. Then, the location information and the direction information of retinal blood vessel skeleton are computed and the modified directional local contrast method is applied to detect the vessel boundary. Last, the width of retinal blood vessel diameter is obtained by computation. The proposed method provides an assistant means for obtaining the blood vessel's morphological information and quantized data.