Research On Carotid And Vertebral Arteries Segmentation In CTA Images Based On Deep Learning

The head and neck CTA images were used to observe the carotid and vertebral arteries.This method is of great significance in the clinical research to assist doctors in the diagnosis of atherosclerosis and stroke,moreover,the segmentation of the carotid and vertebral arteries is one of the key steps.Therefore,the research of segmentation carotid and vertebral arteries in CTA images based on deep learning method is conducive to improve the diagnosis efficiency and reduce the workload of doctors.Firstly,a data preprocessing method for original head and neck CTA images is proposed: in view of the problems of narrow distribution range of carotid and vertebral arteries pixel value and unclear blood vessel boundary in CTA images,the method of truncated normalization is applied.And then the CTA images are adjusted to fit the input size of the network by the way of the image interpolation algorithm.Secondly,the research is taking advantage of V-Net network to segmentate the carotid and vertebral arteries.Per as the characteristics of carotid and vertebral arteries,CVA-Vnet network is proposed,which is improved on the basis of V-Net network as follows: 1)The multi-resolution output fusion in the decoding process of network is beneficial to obtain the deep semantic information of the network,meanwhile,the whole and detailed parts of the carotid and vertebral arteries are taken into account;2)The common convolution in the network is changed into Hybrid Dilated Convolution,and with the enlarging of the receptive field,the feature information of vascular details can be taken better care of;3)Dropout regularization technology is introduced into V-Net network to reduce the risk of over fitting and enhance the learning capacity.Eventually,the predicted label data are spliced and restored to the original size of CTA image through post-processing,and the segmentation results of carotid and vertebral arteries are obtained.The research use 3D head and neck CTA images provided by the project partners for experiments,and 14 sets of CTA image data were randomly selected for testing.Comparing the improved network structure which is proposed in this paper with other different deep learning network structures,the experimental results show that the average value of dice similarity coefficient is 85.74% by carotid and vertebral arteries segmentation algorithm based on deep learning,which is significantly improved compared with other methods,especially in the area of external carotid artery.It shows that the segmentation method of carotid artery and vertebral arteries proposed in this paper has more advantages,which is favourable for assisting doctors in clinical diagnosis and treatments.