Research On CT Image Segmentation Algorithm Of Aortic Dissection Based On Deep Learning

Aortic dissection(AD)is a kind of high-risk aortic vascular disease with high mortality,of which the pathological features are a rupture of the tunica intima of the vessel wall.The rupture let the blood flow between different wall layers and divide the vessels in two parts,creating two lumens called true and false lumens.The main treatment for aortic dissection is endovascular repair,which achieves true cavity remodeling by sealing the dissection of the dissection with a stent.Preoperative diagnosis planning and postoperative treatment of aortic dissection require objective quantitative quantification of structural parameters,precise positioning of the tearing intima and tear position.Segmentation of aortic dissection CT images has important guiding significance for preoperative diagnosis,surgical planning and postoperative repair,but manual segmentation is a time-consuming and laborious work.At the same time,subjectivity will also affect the diagnosis accuracy.Therefore,automatic and accurate segmentation of aortic dissection images has become an important topic in computer aided diagnosis and treatment.In this thesis,a deep learning based aortic dissection segmentation method is studied,which realizes the complete visual display of the aortic region and the tearing intima.The main research work is as follows:1.According to the gray distribution of aortic dissection image,a segment-wise parametric model of aortic dissection image is established.The linear Kalman filter algorithm is used to estimate the positional relationship between aortic segments,and the aorta is completely segmented incrementally.We design a fusion algorithm which is a combination of multi-seed three-dimensional region growing and morphological operations to improve the degree of automation of the algorithm and realize the segmentation of the intima.The experimental results show that the scheme is only suitable for low-precision demand scenarios and lays a foundation for the study of subsequent aortic dissection accurate segmentation algorithms.2.This thesis focuses on the segmentation accuracy of different two-dimensional fully convolutional networks in the aortic region and tearing intima,and analyzes the insufficiency and limitations of the twodimensional fully convolution network,i.e.,they are sensitive to the shape position and unable to utilize the spatial relationship between three-dimensional image layers to guide the learning.And we propose an improvement plan for the above two points.Finally,the validity of the fully convolutional networks and the correctness of the improvement direction are verified by experiments.3.In order to solve the problem of low precision of aortic dissection image,discontinuity of segmentation results and sensitivity to shape and location in the previous study,this thesis design a new scheme for precise segmentation of multi-stage aortic dissection images.Firstly,the centerline is constructed based on the position parameters estimated by the parametric modeling scheme.Then the aortic straightening reconstruction is performed based on the multi-planar reconstruction algorithm to achieve the shape and location normalization of the aorta.Finally,a three-dimensional full convolutional network MPRNet is designed to improve the accuracy and continuity of segmentation results of aortic dissection in three dimensions.This method has a significant advantage over the performance of existing algorithms,achieving precise segmentation of the aortic region and tearing intima.