Research On Retina Image Segmentation Method Based On Fully Convolutional Encoding-Decoding Structure

Ophthalmologists can diagnose whether a patient has certain diseases,such as diabetes,hypertension,and glaucoma,by observing the shape,size,and diameter of blood vessels in the fundus images.However,the existing fundus images cannot clearly feed back corresponding information to the doctor.Therefore,it is necessary to segment the retinal blood vessels of the fundus images.However,due to the inherent complexity around the optic disc,retinal vessel segmentation has always been a medical problem.In recent years,with the advancement of artificial intelligence technology,the automatic fundus image segmentation technology based on deep learning is receiving more and more attention from researchers.In this thesis,the deep learning network model commonly used for retinal blood vessel segmentation was studied in detail,and two innovative structures were proposed.Due to the unsatisfied effect of microvessel image segmentation,this thesis proposed a fully convolutional encoder-decoder based on depthwise separable convolution and channel importance sotring.The network introduced depthwise separable convolution to realize the separation and extraction of channel and spatial features.The importance of each channel was ordered according to the corresponding weight value to identify important channel information and suppress ineffective noise.The model was applied to the retinal vessel segmentation experiment,and the experimental results proved that the proposed method is beneficial to the study of fundus microvascular segmentation.Aiming at the problems of poor segmentation of the fovea disc and unclear vascular segments,this thesis proposed a multi-scale channel importance sorting and important spatial information positioning(MSCS).The model firstly employed the channel importance sorting module to suppress useless feature responses in the encoding process and identify effective channels to better identify capillaries and end retinal vessels.Then,in the decoding stage,the spatial attention mechanism was introduced to extract the position information of the multi-scale feature map to better locate the position of the blood vessel.Finally,in order to reduce the model parameters,improve the operation rate,and obtain a larger receptive field,the multi-scale asymmetric cascade convolution module was designed.Experimental results showed that the proposed method can enhance the clarity of retinal blood vessels while reducing the false positive rate of end blood vessels.