Research On Intelligent Blood Vessel Segmentation Algorithm Based On Image Structure Features

Medical image segmentation is the basic part of medical data processing.Its accurac has a great impact on other follow-up work.Most of the existing studies are based on image segmentation directly,and do not make full use of prior information such as sequence features of scanned images,which makes it difficult to segment small targets in images completely and accurately.In order to segment small blood vessels better,this paper proposes two deep learning segmentation methods based on image structure features.A hybrid Multi Strategy supervised segmentation method is proposed based on the image structure features.When the manual annotation information is sufficient,the supervised method can be well trained.The main improvements of this method are as follows: first,adding the image structure features that can retain the spatial characteristics of organs and tissues in the scanned image as the prior information of segmentation to enhance the segmentation accuracy.Secondly,convolution,MLP and transformer are mixed to extract different levels of feature information in the image,and make full use of the advantages of different methods in different levels of feature extraction,which can effectively extract local features and long-distance features in the image.Thirdly,the attention mechanism is redesigned,which can comprehensively consider the channel characteristics,spatial characteristics and location information to reallocate computing resources to the target area.Fourth,the loss function can dynamically adjust the loss weight according to the area of the segmentation target in the label,and guide the model to improve the segmentation accuracy of small blood vessels.Compared with the existing medical supervised segmentation methods,the proposed supervised method performs vascular segmentation on the lung CT data set collected in this paper.The overall results are in precision,dice,Miou and Hausdorff＿95.The four indexes were improved by 1.7%,1.4%,1.7% and1% respectively,and the segmentation results of ordinary blood vessels and small blood vessels were improved.Experiments show that the algorithm can effectively improve the accuracy of vascular segmentation.In order to reduce the labor cost of labeling data,a multi-stage semi supervised segmentation method based on generating countermeasure network is proposed on the basis of the former method.This method uses the powerful image distribution modeling ability of the generation countermeasure network to obtain the feature information and directly generate the mask.In the training process of the algorithm,only a small amount of labeled data is needed to achieve a good segmentation effect.The main improvements of this method are as follows: first,add image structure features to help the coding network map the image to w space.Second,two generators are used to generate background and blood vessels respectively,and two discriminators are used to resist the image generation ability of the training generator and the alignment ability with the generated mask.Thirdly,the coding network is used to map the image to w space.W space is the middle space of potential space,which can decouple the feature factors.Compared with the existing medical semi supervised segmentation methods,the semi supervised segmentation method based on generative countermeasure network proposed in this paper has the best result of vascular segmentation on the lung CT data set in this paper,in precision,dice,Miou and Hausdorff＿95 and other indicators increased by 1.3%,1.0%,0.9% and 0.6% respectively,and ordinary vessel segmentation and small vessel segmentation on this data set also improved.Experiments show that the semi supervised method proposed in this paper can effectively improve the accuracy of vascular segmentation.Experiments show that the proposed supervised segmentation method can effectively improve the accuracy of vascular segmentation.The semi supervised method proposed in this paper can realize the task of model training and segmentation with only a small amount of labeled data.