Research On Blood Vessel Segmentation Method Of Coronary Angiography Based On Deep Learning

Heart disease is one of the catastrophic diseases that threaten human health,and coronary angiography is an important helper in diagnosing this disease.Coronary angiography Image vascular segmentation is the key to digitizing and standardizing vascular data with complex structures.The accuracy of the segmentation results not only directly determines whether the lesion site can be accurately located,but also assists the doctor in understanding the development of coronary heart disease,and is the basis for a series of medical research including vascular dynamics research.Traditional segmentation methods,including matched filtering methods and region growing methods,require complex pre-processing steps,relying on artificial design features,and require designers to understand relevant medical common sense.In recent years,deep learning,especially deep convolutional neural networks,has achieved significant breakthroughs in computer vision by relying on weight sharing,automatic feature extraction,and computer computing performance.This paper improves the coronary vascular segmentation method using deep learning techniques.The main research work is as follows:(1)For coronary angiography images,the vascular structure is complex and variable,and the contrast agent is unevenly distributed,thus causing artifacts and other noise problems.This paper builds a multi-scale convolutional neural network model using parallelism.This network result corrects local features through global features by inputting images of two different scales,making the segmentation results more accurate.(2)The traditional cross entropy loss function is improved for the impact of the problem of unbalanced sample size when using deep learning.The improved function not only alleviates this problem,but also improves the classification accuracy of difficult samples.(3)When applying migration learning to small sample data,training directly using the sample of the target domain may cause over-fitting or parameter structure damage.This paper proposes a small sample migration learning method by limiting parameter learning,so that the parameters of the target domain are consistent with the parameters of the source domain,and the strategy is solved by FOBOS algorithm.Through comparison experiments,the results are analyzed and compared.It is found that the proposed network model can effectively distinguish artifacts and noise and segment blood vessels more accurately.The improved loss function can effectively improve the results in the training of unbalanced sample size,increase the recognition of difficult samples by the model,and make the network more robust.Using the migration learning strategy to obtain the model for coronary vascular segmentation,the accuracy is better than usual.