Retinal Vessel Segmentation Algorithm Based On Improved CNN Architecture

The detection and analysis of retinal blood vessels in the fundus plays an important role in the screening and diagnosis of many ophthalmic diseases.According to the changes in the diameter,curvature and color of the blood vessels,it can effectively help doctors to judge ophthalmic and internal diseases.Retinal vessel segmentation is a basic step in the clinical examination of ophthalmic diseases,which is helpful for the visualization and quantification of lesions.Therefore,the accuracy of retinal vessel segmentation is very important.The manual segmentation method is time-consuming and laborious,and the topological structure of blood vessels in the fundus image is complex and varied with different shapes.It is difficult to accurately analyze and judge the retinal blood vessels only by manual segmentation method.In contrast,the computer-aided diagnosis system can achieve accurate segmentation of retinal blood vessels,reduce manual burden and promote the efficiency of clinical diagnosis of ophthalmic diseases,so it has great application value and prospects at this stage.To this end,three retinal vessel segmentation algorithms are proposed in this thesis: a U-shaped segmentation algorithm based on multi-scale feature fusion,a U-shaped segmentation algorithm based on adaptive aggregation of feature information,and a lightweight high-resolution network segmentation algorithm.The main research contents are as follows:(1)Considering the different scales of retinal blood vessels and the changeable feature information,etc.Meanwhile,the traditional U-Net model has a single feature extraction method,a U-shaped retinal vessel segmentation algorithm based on multi-scale feature fusion is proposed.Firstly,the ordinary convolution is replaced by the residual module in U-Net to achieve the reuse of features.Then,the parallel multi-branch structure and pyramid pooling block are added at the head,tail and bottom of the network respectively to expand the receptive field of extracted features and promote the fusion of features at different scales.Finally,in order to balance the weight ratio of foreground and background dynamically and eliminate the influence of noise in fundus images,an attention gate mechanism is introduced in the skip connection of the network.(2)In order to segment retinal vessels more accurately and improve the ability of the algorithm to identify microvessels,a U-shaped retinal vessel segmentation algorithm based on adaptive aggregation of feature information is proposed.Firstly,the feature selection module introduced in the encoder part not only strengthens feature transfer,but also enables the network to selectively emphasize feature information.At the same time,a densely connected atrous spatial pyramid pooling module is embedded at the bottom of the network to effectively capture blood vessels at different scales,so that the network can generate richer and denser contextual information.Secondly,an adaptive aggregation module is constructed in the decoder part to aggregate the semantic information in each level of the encoder part and transmit it to the subsequent layers,which is beneficial to the spatial reconstruction of retinal vessels.Thirdly,a joint loss function is introduced to facilitate network training to achieve a more balanced segmentation between vessel pixels and non-vessel pixels.(3)Aiming at the low contrast of fundus retinal images,the complex morphological structure of blood vessels,and the common problems of blurred blood vessel boundaries,obvious noise and inaccurate segmentation of microvessels in the segmentation results of most existing algorithms,this paper proposes a lightweight high-resolution network(LHR-Net)model to improve the segmentation performance of retinal blood vessels.LHR-Net consists of a high-resolution main path,two low-resolution branches,and a multi-scale feature extraction branch.It enhances the representation capability of the network by using before-activation residual block..To generate rich spatial and semantic information in the high-resolution main path and facilitate better prediction of microvessels,a designed parallel channel attention mechanism is embedded into the network.Furthermore,atrous convolutions with different dilation rates are employed to extract multi-scale features of blood vessels.In this thesis,the above three algorithms are verified on the public fundus dataset.Experimental results show that the three segmentation algorithms can segment more fine and complete blood vessel structures with higher sensitivity,and achieve better performance than most existing algorithms.