Research On Retinal OCT Image Segmentation Method For Macular Edema Detectio

Macular Edema(ME)is a disease that occurs in the retina.It usually obscures the vision of the patient,and even leads to blindness in severe cases.ME is most likely to occur in diabetic patients and the elderly.The Inner Limiting Membrane(ILM)to the Retinal Pigment Epithelium(RPE)in the retina is often affected by ME lesions,resulting in morphological changes.Optical Coherence Tomography(OCT)is commonly used in clinical practice to achieve clear imaging of the retinal structure of patients.The effective segmentation of ME and retinal ILM-RPE layer has great reference significance for judging the severity,activity and therapeutic effect of ME.However,manual segmentation has some limitations due to its low efficiency,strong subjectivity and poor repeatability.Therefore,it is urgent to explore a fast and accurate automatic segmentation method for ME and ILM-RPE layer of retinal OCT images.For this task of retinal OCT image segmentation,the deep learning method is adopted in this thesis to design a Residual Convolution U-Net(RCU-Net)network based on residual attention weighting mechanism,which can realize the segmentation of ME and ILM-RPE layer of retinal OCT images.A variety of strategies are designed to achieve accurate segmentation results: Aiming at the problem of insufficient and unbalanced data,a data generation method based on GAN network is designed to extend the retinal OCT data set,so that the generalization ability of the network is improved;Aiming at the problem that ME boundary and retinal layer features are not obvious enough,an improved residual attention weighted structure is proposed,which can effectively strengthen the useful features in channel and space without affecting the gradient propagation while deepening the network,so that the network can better learn different levels of information;Aiming at the optimization of retinal images segmentation network model,the Lovász-softmax loss function is used to guide the iterative adjustment of network parameters.In the training process,the optimal segmentation model can be obtained for the Intersection over Union(Io U)between the segmentation prediction result and the real label,thereby improving the segmentation performance of the network.The experimental results show that the MIo U of the RCU-Net designed in this thesis reaches 88.595% and the accuracy reaches 99.171% for retinal OCT images segmentation,which proves that the proposed method has obvious effectiveness in ME regions and ILM-RPE layer segmentation task of retinal OCT images.