Study On Coronary Artery Segmentation And Centerline Matching In X-ray Angiographic Images

In recent years,the coronary artery disease has been one of the main threats to human health.Although there are many advanced imaging modalities,X-ray angiography is still taken as gold standard for interventional diagnosis of coronary artery disease for its fast imaging speed and high imaging resolution.Computer aided diagnosis in x-ray angiograms can help doctors diagnose and make treatment prescriptions,which is of great importance.However,current angiograms image processing methods(e.g.segmentation,centerline extraction,matching)require tons of interactions,which limits the application of computer aided diagnosis in clinical application.This thesis,by focusing on the limitations of the current image processing methods in angiograms,discusses several key issues: angiograms image matching,coronary segmentation,centerline extraction and matching.Most methods are automatic,which help reducing doctors’ interactions and lay the foundation for other studies like 3D coronary reconstruction and hemodynamic analysis.Detailed work is introduced as following:(1)An angiogram image matching method is proposed based on deep features.First,this dissertation constructs a siamese convolutional neural network to extract the local feature descriptor of the angiogram.To overcome the problem of lacking in training samples,it is firstly trained on samples from natural images,and finetuned on several angiography samples.Secondly,a hierarchical matching strategy is used to generate sparse matches between images.Finally,a dense interpolation method based on the Euclidean distance and feature distance is proposed to generate the dense matches between images.(2)An automatic vessel segmentation method based on multiple-channel convolutional neural network is proposed.Firstly,a pair of live and mask images are matched by our method in(1).Secondly,patches from live images and their corresponding patches from mask images are put into a two-channel CNN,and a coarse segmentation result is achieved.Thirdly,the coarse segmentation result is modified into a region-of-interest area by morphological operations.Finally,pixels in the region-of-interest area are further classified by a multiscale convolutional neural networks to get the final fine segmentation result.Experiment results demonstrate that the proposed method is able to remove most background noise and keep details of the small vascular structures.(3)An automatic centerline extraction and matching method is proposed.Firstly,the segmentation method in(2)is used to get the segmentation result.Secondary,the segmentation result is filtered by multiscale Gaussian filters,and the coarse centerline is extracted by applying non-maxima suppression on the filtered images.Thirdly,shortest path tracking on the connection probability map is employed to search the optimal path of connection around the gaps.Fourthly,the centerline is turned into a graph,and all candidate vessels are found by graph searching using the position information of the vessel in the last frame.Finally,the candidate vessel with minimal dynamic-time-wraping distance with the vessel in the last frame is considered as matched vessel.