Retinal Vessel Segmentation Based On Improved U-Net

Because of its unique visibility,retinal vascular images are used by doctors as an important reference for clinical application.For example,diseases such as diabetes and hypertension will cause changes in the shape of blood vessels,so doctors can judge the symptoms of patients.Therefore,the study of retinal vascular images Play a critical role in the diagnosis and treatment of cardiovascular diseases.Segmentation by doctors is inefficient and occupies medical resources,and manual segmentation has large errors,which is difficult to avoid.Therefore,it is particularly urgent to seek more intelligent automatic segmentation technology.The emergence of computer aided segmentation not only solves the above problems,but also provides guarantee for doctors to make accurate judgment,which has high research value and broad application prospect.To this end,we made the following research:(1)In order to segment retinal blood vessels more accurately,we propose an improved dense residual U-shaped network model to segment retinal blood vessels.Firstly,the soft threshold function and the attention mechanism are integrated into the residual structure,and the unnecessary features are set to 0 by the attention mechanism and the soft threshold function,so as to improve the efficiency and accuracy of segmentation.Secondly,the codec part of U-shaped network is connected with dense residual structure,which strengthens the mutual transmission of different features,makes the information of the network realize the sharing of all layers,and improves the feature recognition ability of the network.(2)In order to improve the segmentation ability of complex images,we propose an improved cyclic residual U-type network model to segment retinal vessels.Firstly,the feature reuse i realized by using the quadratic cycle network,and the network features are enriched without adding network parameters.The feature pooling layer at the bottom of the network fuses the features of the coding end of the network and inputs them into the decoding end together to deepen the relevance of the network features.The gated attention mechanism in jump connection is used to suppress noise interference and improve the sensitivity of network segmentation.Mixed loss speeds up the network training test time and makes the network convergence more stable.(3)In order to solve such problems as poor segmentation effect,data overfitting and unbalanced positive and negative samples in fundus blood vessel segmentation,a U-shaped network model combining convolution structure and transformer is proposed.Firstly,transformer is used to build connections between global features,but local details will be destroyed.Therefore,the feature modeling method formed by combining convolution with transformer is used to extract local features.Multi-head attention in Transformer is used to learn global information and establish global dependencies under different resolutions to the greatest extent.In view of the different thickness and complex direction of blood vessels,an asymmetric feature fusion module is incorporated into the skip connection layer of the network to extract the information of different scales of each layer.The three algorithms proposed in this paper were all used for segmentation experiments in public data sets DRIVE and STARE.Experimental results of each chapter showed that compared with traditional methods,the blood blood of vascular images under the three segmentation methods in this paper was clearer and more complete,with higher evaluation indexes,and had better performance indexes compared with previous algorithms.