Research On Coronary Angiography Image Segmentation Method Based On Full Convolutional Neural Network

With the growth of the urban economy and the aging of the population,cardiovascular disease has become the number one cause of death for human beings.The analysis of coronary vessel geometry is of great importance in the diagnosis of cardiovascular disease.In order to make reliable judgments for clinical applications,it is necessary to extract coronary vascular trees by an accurate and robust segmentation method.This paper presents a set of coronary angiography sequence image segmentation methods based on fully-convolutional neural network.Firstly,the image is preprocessed to improve the quality of the sample.then,a full convolution segmentation network VGG-seg is proposed,which is based on the VGG-16 basic network.The last full connection layer is removed,and deconvolution is carried out after each pooling layer.The deconvolution results of each layer are respectively output and linear summation is carried out to classify each pixel.Define the loss function as adaptive class-balanced cross entropy.The network is initialized with the pre-training parameter on Image Net,and 109 cases of clinical coronary angiography sequence data are used as the training set to train the network,and using 40 coronary angiographic sequence as the test set to test the model.Finally,the conditional random field method was used to refine the blood vessel edge.All the pixels of the coronary angiography image are constructed into a fully connected conditional random field,and the total energy of the model is minimized by minimizing the potential energy and the binary potential energy,taking into account the position and gray of the pixels,thereby achieving the fracture of the blood vessel,and then realizing the segmentation of the coronary sequence image.In order to verify the effectiveness of the segmentation method,this article evaluates the segmentation method on the test set data through two indicators.The first is the evaluation index of the segmentation result,including the Dice coefficient,accuracy,specificity,and sensitivity;The second is the evaluation index of clinical diagnosis results,which is the results of segmentation are refined,and blood vessel diameter is measured.The results obtained in this paper were compared with the labeling results of doctors.In terms of network performance,the Dice coefficient obtained in this paper is 0.89,the accuracy is 98.36%,the sensitivity is 93.36% and the specificity is 98.76%,showing certain advantages compared with the advanced methods proposed by predecessors.In the aspect of clinical evaluation index,Based on the results obtained by the segmentation method in this paper,skeleton line extraction and diameter calculation are carried out,Compared with the diameter of the labeled image,the absolute error is 0.382 and the relative error is 0.112,indicating that the segmentation method in this paper can accurately restore the vascular contour.The method of image segmentation of coronary angiography sequence based on full convolutional neural network proposed in this paper has realized the accurate extraction of vessels in coronary angiography sequence images,providing a certain basis for solving the problem of extraction of coronary angiography tree in clinical practice.