Research On Retinal Artery-Vein Segmentation Based On Dual Attention Network

The artery-vein segmentation of the retinal vessels in the fundus is of great importance in the field of medical image processing and can be used to assist in clinical diagnosis and treatment.The retinal vessels in the human fundus include two important types of vessels:arteries and veins.Medical research has shown that many diseases in the human body,such as hypertension,diabetes,atherosclerosis and other systemic diseases,are prone to ocular complications.The use of fundus photography to obtain clear images of the retinal fundus and the use of computer vision technology to digitally analyze the fundus images will therefore provide an important diagnostic aid to medical practitioners.In this paper,we focus on the artery-vein segmentation of fundus retinal vascular images.The main technical tools are based on the recent emergence of deep learning techniques in neural networks and computer vision processing methods based on image enhancement.The main work of this paper is as follows:Aiming at the insufficient use of context Semantic information and dependence between channels in the existing work of fundus retinal artery-vein segmentation network,as well as the problem of artery-vein misclassification that are prone to occur in artery-vein segmentation.A Dual Attention Network Fusion(DANF)model for artery-vein segmentation is proposed.It applies a CSAM(Spatial and Channel Attention Module)to extract the contextual semantic information features of the image and the dependency features between channels separately,in order to address the problem of insufficient use of contextual semantic information and interchannel dependency in the artery-vein segmentation network of fundus retinal images in existing works.Channel attention and Spatial attention work in parallel to efficiently extract the curve structure of the retinal vessels in the fundus and achieve better image segmentation performance.For the misclassification of arteries and veins that is prone to occur in retinal artery-vein segmentation,that is,mistakenly classifying arteries as veins or mistakenly classifying veins as arteries,a network fusion method is used to fuse the separate segmentation of arteries and veins with the results of simultaneous segmentation,and to enhance the results of artery-vein segmentation to reduce misclassification.The DANF model uses Deep Res UNet as the backbone network,which allows more layers of the neural network to fully learn the representation and helps to further improve the segmentation results of the model.Experimental results on three public datasets of retinal fundus images show that the DANF model proposed in this paper achieves significantly improved performance in artery-vein segmentation relative to other baseline methods.To overcome the limited number of open standard datasets for retinal image segmentation currently used in the research area of this thesis,and the small number of training samples in them,as well as the low contrast of blood vessel pixels in the fundus retinal images in the dataset,the high amount of noise information,and the lack of clarity of fine blood vessel pixels.From the perspective of the dataset,an image enhancement approach oriented to the artery-vein segmentation of retinal images is investigated,and the enhanced images are fed together into the DANF model.The experimental results show that on three publicly available datasets of retinal fundus images,i.e.DRIVE-AV,HRF-AV and LES-AV,the structure of retinal vessels is enhanced by applying the image data enhancement approach,and the overall artery-vein segmentation performance is further improved.