Carotid Artery Intelligent Segmentation And Three Dimensional Reconstruction Based On Small Sample Ultrasound Images

Atherosclerosis is the inherent cause of cardiovascular and cerebrovascular diseases.Clinically,the lumen-intima boundary and media-adventitia boundary of the intima-media and the media-adventitia boundary of the carotid artery ultrasound image are manually segmented,and finally the intima-media thickness is measured to determine atherosclerosis and plaque.The condition of atherosclerosis and plaque is time-consuming,laborious and subjective,and the rate of missed diagnosis and misdiagnosis is high.Based on the 3D ultrasound image of the carotid artery,this paper proposes a 2.5D-VNet convolutional network.Based on the optimization of V-Net,our network splits the 3D convolution kernel into separate threedimensional convolution and two-dimensional convolution,which is 2.5D convolution,and the final total network parameter is only 26.3% of V-Net.At the same time,because of the limitations of two-dimensional ultrasound images in plaque recognition,this paper designs a set of automatic acquisition system,which can automatically calculate the heart rate of the collected person,and collect each frame of two-dimensional ultrasound in each carotid artery contraction phase.Our system can obtain accurate and reliable three-dimensional images of the carotid artery.This paper uses a three-dimensional ultrasound acquisition system to collect 100 small samples of carotid artery three-dimensional ultrasound images.The data volume is increased by data enhancement and data expansion then fed to the network training.The segmentation results of the intima-media boundary can be obtained next.What’s more,this paper automatically recognizes and segment the patches according to the patch determination algorithm.Finally,the three-dimensional image segmentation results and the plaque segmentation results are reconstructed in three dimensions,and the inner media and plaque models are obtained.The image segmentation results show that the average Dice,IOU,recall and accuracy coefficients of the 2.5D-VNet network on the mediaadventitia boundary segmentation are 0.854,0.745,0.887 and 0.824,respectively.The average test time for each image is 2.751s;the correlation coefficients on the segmentation of the lumen-intima boundary are 0.827,0.705,0.875 and 0.784,respectively.The average test time for each image is 2.560 s.The segmentation result is equivalent to V-Net and the test time is greatly reduced.The results of plaque recognition show that the intersection ratio coefficients of automatic plaque recognition algorithm for plaque recognition at the internal and external carotid arteries and common carotid arteries can reach 0.839 and 0.785,respectively.The three-dimensional reconstruction model shows that doctors can more intuitively and effectively observe the surface biomorphism and relative position of the carotid artery intima and plaque based on the threedimensional model,which provides new clinical screening for carotid atherosclerosis and plaque auxiliary diagnosis method.