Graph-based 3D vascular segmentation and its applications
Ning Zhu · 2014
In recent years, vessel disease, which is one of the major causes of death around the world, has become an important health problem. Vessel segmentation is an important technique, and it can help the diagnosis, visualization, treatment and surgery planning for vessel diseases. In this thesis, we first review existing techniques for 3D vessel segmentation. Then several vessel segmentation techniques are proposed. An efficient centerline extraction method (Minimum Average-cost Path (MACP)) is firstly proposed. The traditional issue of minimal path methods, namely shortcut problem, can be solved with MACP. Then based on the centerline obtained with MACP, two centerline-based segmentation methods, Optimal Cross Sections and Graph Optimization via Graph Cuts are then proposed for obtaining accurate vessel cross sections and vessel segmentation, respectively. Though centerline-based method is popular and efficient, the performance of 3D vessel segmentation heavily relies on the accuracy of centerlines. To solve this problem, segmentation methods without ROI restriction are proposed. Random Walks with Adaptive Cylinder Flux (ACF) based Connectivity, Power-watershed based Optimization with Tubularity Markov Tree (TMT) are two methods that do not rely on any centerline input. The Adaptive Cylinder Flux and Tubularity Markov Tree are models with tubular feature detectors which can help the detection of the vessel structure during the segmentation. Random Walks based Calcium Elimination and Distal Vessel Augmentation (CEDA) Optimization, and Single Target Segmentation with Adaptively Pruned Tubularity Markov Tree model (AP-TMT) are two extensions. The evaluation results of Random Walks based CEDA Optimization on public evaluation framework have demonstrated that the method is more accurate than all state-of-the-art methods on segmentation accuracy for healthy vessels. For vessel segments with diseases, comparable results are obtained. Vessel stenosis is one major type of vascular diseases. Because of the complex structure of stenoses, both detection and quantification of stenoses are challenging. In this thesis, two segmentation based methods (PWIS and type-based) and one learning based stenosis detection method are proposed and evaluated on public evaluation framework for stenosis detection. Type-based method achieve higher sensitivity than all existing methods. The overall performance (average rank) of learning based method is better than all existing methods.