Research On Coronary Artery Centerline Extraction Method Based On Deep Tracking Network

It is of great significance for clinical practice to obtain reliable coronary centerline from CT angiography data effectively,and it may provide preconditions for the assessment of coronary stenosis and atherosclerotic plaque.Therefore,researchers began using various methods to try to extract the centerline from computed tomography angiography(CTA)data.However,there are still many difficulties in the extraction of the coronary centerline due to its small size and complex structure,low-dose imaging noise and reconstruction artifacts caused by breathing and heartbeat,the existence of these problems hinders the extraction of coronary artery centerline.To solve the above problems,in this paper,a multi-task coronary artery centerline extraction method based on deep tracking network is proposed.The contribution includes two categories:(1)A Patch-based Iterative Network(PIN)for fast and accurate coronary initial point detection method.The location of seed points is particularly important in the extraction of coronary artery centerline,so in this paper,the seed points are automatically acquired through the network and they only include the left and right coronary ostium.Considering the left and right coronary ostium of the coronary centerline as anatomical landmarks,by using the convolutional neural network to learn the image features around the landmarks,so as that the network can automatically obtain the left and right coronary ostium in unknown images.The obtained left and right coronary ostium are used as seed points for subsequent centerline extraction.It is proved by a large number of experiments that,the starting points detection method of coronary artery centerline proposed in this paper is feasible,and the left and right coronary ostium finally obtained can be used as the initial points for subsequent centerline extraction.The test takes only 3.53 s,and the average distance error is 5.50±0.31 pixel.(2)Multi-task coronary artery centerline extraction method based on deep tracking network is used to extract the centerline of coronary artery.In many current coronary centerline extraction algorithms,only the feature of the image block where the current centerline point is located is used,and the sequence information of the image block where the continuous centerline points are located is not mined.As a result,the current methods often fail to track the coronary artery centerline in some position such as poor coronary angiography or coronary lesion.In addition,the current methods for extracting the centerline of coronary arteries are all based on a large number of seed points,there are seed points on the coronary branched,so the bifurcation of the centerline does not need to be considered.Therefore,we designs a using only two initial seed points method,which can not only predict the direction of the centerline but also judge the bifurcation points at the same time.It can not only learn the features in the current image,but also make good use of the sequence information of the image blocks.The two initial points obtained in the first step are used to achieve the tracking of centerline,and the automatic extraction of the centerline of the coronary artery is finally realized.This paper uses the evaluation method in the Rotterdam Coronary artery algorithm evaluation framework.The result shows that OV,OF,OT and AI is 90.4%,78.1%,92.2% and 0.25 mm.It shows tha the method proposed in this paper has a good performance in automatic extraction of coronary artery centerline.