Cerebrovascular Segmentation Algorithm Based On Multi-scale 3DUnet

In recent years,the incidence and mortality of stroke caused by cerebrovascular diseases in China have risen year by year,which is vitally interrelated with the increasingly serious aging phenomenon in China.This kind of disease has become the primary threat to a large number of lives and health.Vascular morphology(the shape,location,distribution,and thickness changes of blood vessels)is the gold standard for determining the cause and type of brain diseases,and is also an important reference for diagnosis,treatment,and navigation of neurointerventional surgery.As the basis and premise of the above,considerable researchers and scholars have been attracted by cerebrovascular segmentation technology and took it as their research goal.In a variety of medical images,CT angiography(CTA)based on ultrasound imaging is commonly used for screening because of its advantages of non-invasive and low-cost.However,it can lead to a serious imbalance between the target foreground and background in brain vascular segmentation with small diameter,multiple branches,and complex structure due to the wide threshold range of CTA images,seriously affecting the speed and accuracy of CTA brain vascular segmentation.In order to improve the accuracy and speed of cerebral vascular segmentation,this paper analyzes the practical difficulties of traditional segmentation algorithms,and designs a high-precision and robust cerebral vascular segmentation algorithm for CTA cerebral vascular images based on deep learning.This algorithm solves the problems encountered in existing segmentation networks such as missing context information,low automation,high model complexity,and limited small samples.The specific content is as follows:First of all,a context based 3DUnet algorithm for cerebral vascular segmentation is proposed in this paper.Faced with the problem of shallow information loss in multi-layer convolution,this paper designs a context module to fuse the semantic feature information extracted from deep convolution with the spatial information extracted from shallow convolution to achieve information reuse.Adding a context module to the 3DNet network,integrating encoder output features and initial input features through jump links,and using channel attention to filter important features can be propitious to improve the segmentation performance of the algorithm for small blood vessels in the brain.For the sake of achieving better performance of the algorithm under hardware constraints,this paper designs to use GN(Group Normalization)for normalization operations,and proves the rationality and effectiveness of the model algorithm through experiments.In addition,taking the accurate segmentation of cerebral vessels as the research goal,this paper proposes a multi-scale 3DUnet based segmentation algorithm for cerebral blood vessels based on the above algorithms.A multi scale jump link module based on the context module is designed to achieve multi scale feature learning at different levels.By fusing each layer of encoder output with decoder features through dense jump links and redistributing the weight through channel attention,more comprehensive feature information of cerebral blood vessels can be obtained,precise segmentation of small cerebral blood vessels can be achieved,and algorithm performance can be improved.In order to obtain richer semantic features from the segmentation results,this paper designs an output feature fusion module,which aggregates and filters the output of each layer of decoder to obtain the final segmentation results.Experimental results show that the context module,multi scale jump link module,and output feature fusion module designed in this paper have significantly improved performance,and the combined evaluation index of the three parts is optimal.