Research On Image Segmentation And Stenosis Detection Of Coronary Angiography Based On Deep Learning

Coronary artery disease is the most common cardiovascular disease and one of the most common causes of death in the world.X-ray coronary angiography is the main imaging technique used to diagnose and treat coronary artery disease.Segmenting blood vessels and detecting stenotic vessels from coronary angiography images can help doctors diagnose diseases accurately.At present,although many methods have been proposed by researchers,the accuracy of coronary vessel segmentation and stenosis detection remains a challenge.In this context,this paper has made the following research.Firstly,the dataset of coronary angiography images was established.The coronary angiography data from real cases were collected,and after desensitization,the data were sorted and classified according to different angiographic positions and the degree of vascular stenosis,and a set of reasonable data labeling scheme was developed.Secondly,vascular segmentation experiments are carried out by using traditional segmentation methods and deep learning methods.The traditional segmentation method uses the Otsu threshold segmentation algorithm and improves it.The improved Otsu algorithm can segment a clear trunk of blood vessels.The deep learning method uses UNet and Swin-Unet.The UNet is optimized so that the size of output image is equal to the size of input image.When training Swin-Unet,a loss function based on penalty is used to improve the effect of vascular edge segmentation.Because the Swin-Unet uses Swin Transformer structure and penalty-based loss function,the segmentation effect is the best.Finally,three different convolution networks are used as feature extraction networks of Faster R-CNN to detect stenotic vessels,and the location of stenotic vessels is successfully detected and the probability of vascular stenosis is predicted.The experimental results show that the detection effect of stenotic vessels using Resnet50 as feature extraction network is the best,and the mAP of verification dataset is finally stable at 63%.In order to accurately quantify blood vessels,a quantitative measurement method for measuring vascular parameters is proposed,and a measurement software is designed.