Research On Refraction Vessel Segmentation Based On Multiscale Feature Fusion

Cardiovascular disease,diabetes can cause changes in blood vessel morphology.Analysis of retinal blood vessel morphology can make early diagnosis of related diseases.Therefore,designing a high-performance retinal vessel segmentation method has great research significance and application value.This article first analyzes the results of retinal vascular segmentation using the U_Net model,and makes in-depth research on several problems in this model in retinal vascular segmentation,and proposes the improvements and improvements made in this article in response to these problems.In this paper,each improvement point is compared with the U_Net model to verify the feasibility of the improvement points.Finally,the improvement schemes of the small points proposed in this paper are fused to obtain the final model structure,and the final model is analyzed for performance.The specific research content is as follows:(1)Aiming at the problem of inaccurate segmentation of fine blood vessels with the result of U_Net model segmentation,this paper proposes to use a multi-scale convolution kernel to extract features from the original image,which constitutes the Inception convolution module designed in this paper.The feature maps obtained through this structure have receptive fields of different sizes,so it can have a better detection effect on vascular structures of different sizes in retinal vascular images.(2)In order to enhance the outline of the retinal vessels in the model segmentation results,the outline is clearer.In this article,the structure of the vascular position features obtained by upsampling using the maximum index,the deep semantic features obtained by deconvolution and the shallow contour features are designed in the upsampling structure.This makes the model more efficient in identifying vascular structures.(3)This paper uses a combination of DICE loss and cross entropy loss in the design of the loss function.The DICE loss is the similarity loss between the model prediction result and the target result,and the cross entropy is the loss of the class probability of each pixel in the model prediction result.A cost-sensitive matrix method is added to the cross-entropy loss to solve the problem of the imbalance between the number of blood vessel pixels and the number of background pixels in the image.The accuracy of the model designed in this paper on the DRIVE dataset can reach 0.9694,the sensitivity can reach 0.7762,and the specificity can reach 0.9835.On the STARE dataset,the accuracy can reach 0.9537,the sensitivity can reach 0.7721,and the specificity can reach 0.9885.The effectiveness of the model designed in this paper.