SD-OCT Diabetic Retinopathy Automatic Segmentation And Relationship Analysis

Diabetic retinopathy(DR),a highly specific vascular complication caused by diabetes,has become a leading cause of vision loss and even blindness.In recent years,the spectral domain optical coherence tomography(SD-OCT)has been widely used in the field of ophthalmology,especially in the diagnosis of retinal diseases,because it can clearly show the multi-layer structure of retina and small lesions.In this paper,the image processing and analysis is adopted to study the DR in SD-OCT retinal images and the main research contents are as follows:(1)An automatic segmentation method based on self-adaption threshold and region growing for hyper-reflective foci(HRF)is proposed.Firstly,the layer segmentation method is used to limit the target region.Then we obtain the seed point image through the self-adaption threshold method.Finally,the region growing method based on human vision characteristics is adopted to segment the targets.Experimental results demonstrate that the proposed method can accurately segment most of the HRF in SD-OCT images.(2)An automatic segmentation method based on deep convolutional neural network for HRF is proposed.We modify the google and residual networks which perform excellently in image classification to make them suitable for small image patches.A specific data set is built to train the network model according to the distribution and geometric characteristics of the HRF in SD-OCT images.A series of comparison experiments based on different networks and input scales are conducted to acquire the optimal results.Experimental results demonstrate that the proposed method can accurately segment the targets and the results are closer to the manual results visually.(3)Analysis of the relation between HRF and hard exudates(HEs)is performed in SD-OCT DR images.Firstly,the SD-OCT projection image is registered to the color fundus photography(CFP)image and the HEs in the SD-OCT B-scan images can be determined according to those in the registered cropped CFP image.Then five quantitative image features including area,amount,intensity,height and distance to fovea are extracted to investigate the correlation of HRF and HEs,and a disease evaluation model is built based on these lesion features.Experimental results demonstrate the positive correlations in area and amount between HRF and HEs at different stages of DR,with statistical significance(p<0.05).The area and amount can be taken as potential discriminant indicators of the severity of DR.