Coronary Artery Segmentation In CT Images Based On Prior Knowledge

Accurate segmentation of coronary artery lumen on Coronary CT Angiography(CCTA)image plays a very important role in clinical diagnosis of coronary artery diseases.The complete coronary lumen segmented by coronary artery segmentation algorithm will effectively and noninvasively detect functional coronary artery stenosis combined with CT-FFR technology,so as to help doctors plan treatment.In recent years,with the development of deep learning,coronary artery segmentation methods based on deep learning have been proposed,which greatly shortens the coronary artery segmentation time.However,due to the difficulty labeling and the thin and complex structure of coronary artery,the existing fully supervised deep learning coronary artery segmentation algorithms are challenging to segment coronary artery efficiently with a small number of labels.This work mainly studies the coronary artery segmentation algorithm based on prior knowledge.It uses the global and local prior knowledge to assist the segmentation of CCTA images,improves the accuracy of coronary artery lumen segmentation,and reduces the demand for labels.Based on the global prior knowledge of the coronary artery centerline,this work proposes an Examinee-Examiner Network(EE-Net).The global continuity topology prior knowledge obtained by EE-Net is used to constrain coronary artery segmentation and optimizes the output through back-propagation.Finally,the network can get better segmentation results in hard-tosegmented areas.The Drop Output Layer(DO layer)used in the network is used to dynamically balance the categories.The well-segmented areas are discarded in the output characteristic map and the hard-to-segmented areas are weighted so that the hard-to-segmented areas(such as coronary stenosis areas)can get more training opportunities.The plug-and-play network structure can also reduce the consumption of computing resources.The experimental results show that EE-Net has a Dice coefficient of 78.3%,which is significantly higher than other comparison methods,and the segmentation results have a continuous topology,which has the potential for clinical application.From the perspective of local a prior feature,this work proposes a label-efficient coronary artery segmentation framework,Learn and Transfer Network(LT-Net).LT-Net transfers the prior features of tubular structure learned in the vessel enhancement algorithm to the network and guides the network to obtain the ability to perceive tubular structure during the learning process.Finally,it can efficiently guide the network convergence and realize label efficient coronary artery segmentation with small number of labels.Experiments show that compared with other methods,LT-Net pays more stable attention to the features of tubular structure at feature extraction.At the same time,this attention will more effectively guide network learning and make the network converge faster.