Detection And Segmentation Of Central Serous Chorioretinopathy Based On SD-OCT Images

Central Serous Chorioretinopathy,CSC for short,is an idiopathic detachment of the posterior neurosensory retina,which has become a main cause of visual acuity in middleaged men.SD-OCT images of CSC can not only show the location of the lesion area,but also visually display the size and the type of the lesion area,which can help the doctor to diagnose and treat CSC in clinical practice.Therefore,the research and quantitative analysis of SD-OCT images of CSC is of great significance.In this paper,we use image processing and analysis methods to segment the regions of CSC and compare them with other algorithms.The following research contents are included:(1)A method for the segmentation of SD-OCT CSC NRD region based on region restricted three-dimensional region growing is proposed.Firstly,the position of ILM boundary IS/OS layer and lower boundary of RPE are obtained by layer segmentation method.Adaptive threshold segmentation is performed between IS/OS boundary and RPE lower boundary to obtain candidate seeds for 3D region growing.The three-dimensional limitation of the lesion area is obtained by segmenting the restricted projection image between IS/OS layer and the lower boundary of the RPE.Then K-means clustering is performed on the candidate seeds in the restricted area to obtain the seed point.Finally,threedimensional region growing is performed in the restricted area,and the initial NRD region segmentation result is obtained.After some post-processing methods,we obtain final segmentation results.The experimental results show that the method can accurately segment NRD lesions.Compared with the other four methods,it not only has higher segmentation accuracy,but also consumes less time,which is suitable for clinical diagnosis.(2)A CSC segmentation method based on random forest is proposed.First,a layer segmentation method is used to obtain the ILM boundary and the BM boundary,and the range of the region is limited.Then,a 49-dimensional feature vector is constructed for each pixel in the restricted region.Each pixel in the restricted region is classified by random forest classifier.We obtain the initial segmentation result of CSC.With the prior knowledge of the lesion region,we fix some false segmentation.Finally,according to the average intensity difference between the upper and lower areas of PED and NRD forms,the lesions are classified.And the segmentation results of the lesions are obtained.The experimental results show that the method can accurately segment the CSC lesions,and can classify the lesions with strong robustness and high accuracy.