The Research On Retinal Layer Segmentation Of OCT Image

OCT fundus tomography is becoming a hot topic in the field of fundus image research,and the importance of eye fundus diagnosis and treatment is becoming more and more important.With the deepening of the degree of aging in our country and the lack of common eye-protection sense in general,the number of elderly people suffering from eye diseases in recent years is increasing rapidly.Eyes are an important organ of human body.Meanwhile,fundus disease is an early proof of some systemic disease,so the development of fundus image processing technology is of great significance to national health level.Retinal layer segmentation is an important part of OCT fundus image processing,and is the main direction of the research.The retinal image layer segmentation in this paper is divided into image preprocessing and image segmentation.In the image preprocessing work,the conventional method of pretreatment of retinal image is firstly introduced.In order to highlight the hierarchical feature of the retina and reduce the unrelated interference,this paper proposes an improved retinal dual flattening method based on the common methods,the location of the top marginal layer and bottom marginal layer of the retina are determined at the same time.Based on the results of flattening,the method of vascular extraction is proposed to obtain the exact location of blood vessels,and the two marginal layers limit the next processing area.In the image segmentation work,in order to improve the edge coincidence degree,avoid the problem of noise sensitivity and the computational complexity based on pixel segmentation,we use superpixel to initialize the over-segmented retinal image,and proposed an adaptive scale selection based on the thickness of the preprocessed retina.In the process of merging oversegmented regions,we introduce the dynamic region merging method based on local optimal similarity relation and region consistency in detail.Based on the method,this paper propose a new similarity calculation method based on statistical and orientation information.Aiming to avoid the problem of blocking horizontal merging process by interlayer noise points,we propose a new method to eliminate interlayer noise based on regional consistency,which makes the merging result of the same layer more complete.Then,we optimize the merge result to avoid the occurrence of isolated regions.Finally,the edge position in the straightening retina is inverse transformed to find the edge position in the original image,and the whole process of retinal layer segmentation is completed.The experimental results show that the method of straightening and blood vessel extraction have accurate results in different retinal images.The effect of similarity calculating method based on statistical and orientation information is improved by the combining accuracy and the time efficiency of the combination.The interlayer noise is removed effectively to avoid the occurrence of the "separation" situation.Overall,the method proposed in this paper have a better result on retinal layer segmentation.The coincidence degree and time efficiency are higher than the pixel-based segmentation method,and more robust to noise.