Research On 3D Segmentation Algorithm Of Coronary Artery

With the increasing demand for diagnosis of coronary heart disease,medical image-based auxiliary diagnosis has become the main method to improve the diagnostic efficiency.3D segmentation of coronary artery based on CTA data is a key technology for the diagnosis of coronary heart disease,and has become a research hotspot.Due to the factors such as different devices,different patients and uneven distribution of contrast agents,the clarity of coronary arteries in CTA data is different,and 3D segmentation of coronary arteries has become a challenging task.In this paper,the coronary artery segmentation algorithm for CTA data is carried out and improved from three aspects:accuracy,running speed and autonomy.The main contributions are as follows:(1)Proposed a coronary artery segmentation algorithm based on multi-feature region growing.Considering the different scale tubular characteristics of coronary artery,the similarity characteristics of blood vessels under different scales are extracted by the tubular-like filter,and the density characteristics of coronary artery are used together as the constraint of region growing.The algorithm solves the under-segmentation phenomenon that is easy to occur in the region growing algorithm based on single feature,and improves the accuracy of the segmentation algorithm.In the case of a priori threshold,the test Dice coefficient is 0.88,which is 0.2 higher than the single feature method's,while in the absence of the a priori threshold,the test Dice coefficient is 0.6(also higher than the single feature method's),and the algorithm running time is 460 seconds.(2)Proposed a coronary artery segmentation algorithm combining 3D full convolutional neural network and seed point search strategy method based on growing.The proposed segmentation algorithm is improved from three aspects:starting seed point localization,coronary artery block segmentation and seed point search strategy:· Starting seed point localization:According to connectivity of the ascending aorta and coronary artery,the coronary artery root node is automatically located by segmenting the ascending aorta,and provides conditions for automatic segmentation of the coronary artery;·Segmentation of local block data:A 3D segmentation network with multiple receptive field fusion is proposed.In the netwrok,64×64×64 local data and Dice loss function are used to solve the problem of extremely unbalanced label distribution appearing in 3D segmentation of coronary artery.In the design of network structure,pyramid pooling is added at the end of the encoder.The structure of multi-receptive field fusion combines the global information and local information in the data block,which restrains the interference of other organizations well.In the encoder-decoder full convolutional network structure,the Residual module,Batch Normalization,and Dropout operations are added to suppress network overfitting and ensure fast convergence of the network;· Seed point search strategy:A seed point search strategy based on growing thought is proposed.By analyzing the segmentation result of the last block,the center of the next selected block is determined,and the block near the coronary artery is extracted.The strategy is efficient to use the segmentation network;(3)At the end of this paper,according to the segmentation results,the measurement of coronary artery is preliminarily realized by straightening,thinning,segmenting,the cross-sectional area calculation.This lays a foundation for subsequent detection of coronary artery stenosis.In summary,the proposed coronary artery segmentation algorithm in this paper increases the Dice coefficient on the test dataset from 60%to 76%,decreases the running time from 460 seconds to 40 seconds,and achieves automatic segmentation of the coronary artery.The measurement of vascular stenosis is preliminarily realized based on the segmentation algorithm in this paper,which proves the effectiveness of the proposed algorithm.