Research On Layering Method Of Retina OCT Image Based On DPNet

Retinal disease is a common ophthalmological disease,which seriously endangers people’s health.Clinical studies have shown that the occurrence of retinal diseases will cause changes in the thickness of the retina,and even the loss of a certain retinal layer in severe cases.Optical coherence tomography(OCT)imaging technology can clearly display multiple retinal cell layers and some biomarkers of retinal diseases.Accurately segment the retinal layer in OCT images can assist doctors in the diagnosis and screening of various diseases.Due to the low contrast of the boundary of the retinal layer,the disease may also cause the deformation of the retinal layer.Therefore,the traditional segmentation method has a large error,and the retinal layer segmentation still faces a huge challenge.With the goal of accurately segmenting the retinal layer,this thesis deeply researches the retinal OCT image layering method based on deep learning.In order to achieve the precise segmentation of the retinal layer in the retinal OCT image,this thesis designs a semantic segmentation network DPNet that combines the dual attention mechanism(DA)and the pyramid pooling module(PPM).In view of the low contrast of retinal layer boundary in retinal OCT images and the deformation of retinal layer caused by pathological changes,DPNet first uses a dual attention module to reweight the features extracted by the basic network,so that the network could pay more attention to the important features at the retinal layer boundary.Then this thesis uses pyramid pooling to perform multi-scale processing on the weighted features to obtain richer global information.Finally,feature a fusion of different high and low levels of information to enhance the semantic relevance of the context.The dual attention module used in the DPNet network is composed of a position attention module and a channel attention module in a dual form.The position attention module selectively aggregates the features of each location according to the similarity between all location features,and the channel attention module selectively emphasizes the existence of the interdependent channel mapping by integrating the relevant features of all channels.And the dual attention module aggregates the features output of the two modules to enhance feature representation.The experimental results show that the DPNet model designed in this thesis can effectively segment the retinal layer in the normal retinal OCT image and the retinal OCT image with AMD disease,and it also has a better segmentation effect of the retinal layer with low contrast and deformation.A high accuracy rate is obtained on the data set produced in this article,and the MIo U value can reach 87.44%.In addition,this thesis also verified the effectiveness of the added dual attention module and pyramid pooling module.