Research Of Automatic Geographic Atrophy Area Segmentation Algorithm Based On SD-OCT Retinal Image

Visual impairment is an important factor affecting people’s quality of life and there are several of causes of visual impairment.Retinopathy is one of the most important cause of visual impairment,such as glaucoma,diabetic retinopathy(DR) and age-related macular disease(AMD).Recently,with the development of imaging technology,spectral domain optical coherence tomography(SD-OCT) has been widely used in the diagnosis and treatment of the retinal diseases.SD-OCT imaging has the advantages of high resolution,non-contact and fast imaging speed.It can observe the fine three-dimensional structure of the retina.It is of great clinical significance and research value to identify and quantify the lesion areas automatically in SD-OCT retina images by the way of image processing.Meanwhile,as SD-OCT can provide more physiological manifestations of eye diseases,using computer to segment and quantify lesions has become a hot topic.This thesis focuses on the automatic segmentation and quantification of GA in SD-OCT retinopathy images.The main contributions of this thesis are summarized as follows:(1)An automatic segmentation model for GA lesions based on two-stage learning was proposed.Using the axial data of SD-OCT images as samples to construct the segmentation model including off-learning model and self-learning model based on stack sparse auto-encoder.The off-learning model is a general model,which can roughly classified by learning common features.However,due to the diversity of patients and the complexity of retinopathy,the general model is robust to all cases.Therefore,the self-learning model was proposed to consider the characteristics of patients and learn the discriminative features.Finally,the false positives can be reduced effectively and the segmentation accuracy can be improved by fusing the results of off-learning model and self-learning model.(2)An automatic segmentation model for GA lesions based on multi-path 3D three-dimensional convolution neural network was proposed.It will cause the holes of segmentation results because the spatial information of images is neglected when only using the single axial data as samples.Based on the problem mentioned above,a multi-path three-dimensional convolutional neural network was designed.In order to preserve the spatial information of the images,the images were divided into three-dimensional image blocks to provide reasonable input data for the network.Then,fusing the different features acquired by different size of convolution kernels.Finally the softmax loss and center loss were used to supervise the training process of the model to solve the problem of segmentation holes effectively.(3)An automatic segmentation model for GA lesion based on patient-independent was proposed.Influenced by the diversity of patients,the pattern of manifestation of various patient is different.In addition,the images contain optic nerve head and other lesions except GA,which makes the segmentation task more challenging.Based on these factors,samples augmentation and data sets fusion was first adopted to overcome the problem of insufficient samples.Besides,a multi-loss convolution neural network was proposed to monitor the network by fusing the block loss,and the center loss and the softmax loss to establish a more robust automatic segmentation model.