Application Of Deep Learning In Segmentation Of Arterial Vessels In Human Brain CTA

Cerebrovascular disease is one of the important reasons that affect the health of the people.In severe cases,it can lead to accidents such as stroke.Because stroke has the characteristics of acute onset and high mortality and disability rate,how to improve the diagnostic accuracy of cerebrovascular diseases and shorten the diagnosis time in the emergency department is a major problem that needs to be solved in clinical practice.The preferred imaging method in emergency department is Computed Tomography Angiography(CTA)with fast,clear and non-invasive imaging.However,CTA is a three-dimensional image,and its two-dimensional slices need to be checked back and forth during diagnosis,which greatly reduces the efficiency of doctors’ diagnosis.Efficient and accurate computer vision algorithms have brought solutions to this problem:the combination of computer vision algorithms and CTA as a clinical auxiliary diagnostic tool can greatly improve the diagnostic efficiency and accuracy of doctors.In this paper,a set of segmentation system of intracranial arteries is designed according to clinical needs.Main tasks as follows:(1)Annotate the data set.Due to the lack of open source data sets for CTA brain vessel segmentation,this paper collected 77 head and neck CTA images from clinically,used a combination of computer annotation and manual correction to annotate the intracranial arteries,and plans to make them public.(2)Design a strong benchmark algorithm for blood vessel segmentation.In this paper,on the basis of 3D UNet,effective modules such as preprocessing,atrous spatial pyramid pooling,group normalization,and post-processing have been added,and reaches83.51% of the Dice coefficient on the test set of 22 CTA images,which is 6.38% higher than 3D UNet,and uses a visualization method to analyze the role of each module.(3)Propose a multi-task network and loss function to optimize the topological structure of the blood vessel segmentation result.A multi-task network is used to perform blood vessel segmentation and centerline extraction at the same time,and a loss function based on the centerline to constrain the vessel topology structure is proposed,which realizes the optimization of the vessel topology structure during segmentation.The traditional segmentation evaluation criteria and the Betti number in algebraic topology are used to evaluate the performance of the algorithm in traditional segmentation tasks and the ability to maintain the topological structure of blood vessels.(4)Develop a software with interface.The software has a built-in segmentation algorithm designed in this article,and provides easy-to-operate functions such as image viewing,algorithm prediction,and manual correction,etc.,to help improve the accuracy and efficiency of clinical diagnosis.In this paper,a complete CTA cerebral artery segmentation system is built from the aspects of data collection and annotation,algorithm design and improvement,software and interface design.The system can help doctors improve the efficiency and accuracy of diagnosis,has a certain practical value,and also provides a certain reference significance for the design of the future cerebrovascular disease diagnosis platform.