| Stimulated by the goals of improving detection of low contrast and narrowvessels, and eliminating false detections at nonvascular structures, a new methodis proposed for extracting vessels in thoracic3D CT images. This method iscomposed of the segmentation of the lung, the extraction of the vessel and thetracing of the vessel tree. And no manual intervention is needed in theprocessing.The core of the method is a new decision criterion that combines thecontinuity in vessel direction, the existence of the cross section, matched-filterresponses, confidence measures and vessel boundary measures. Only voxelsmeeting the continuity in vessel direction and the existence of the cross sectionin one (or more) of52directions will be preserved. These two features canremove noise such as discs with certain thickness and irregular entities. Inaddition to this, the segmentation of the lung can narrow down the processingscope which means some noise is removed in this step.The existence of cross section is used to check whether the cross section isapproximate to circle, eclipse or other defined shapes. Matched filter responsesare derived in multi-scale space to extract vessels with widely varying widths. Avessel confidence measure is defined as a projection of a vector formed fromnormalized pixel neighborhoods into a normalized vector of ideal vessel profile.Vessel boundary measures and associated confidence measures are computed atpotential vessel boundaries. These responses form a10-dimensionalmeasurement vector at each voxel. A schema is used to develop a mapping ofthis vector to background or vessel at each voxel. Finally, the vesselnessdecision criterion is embedded into a vessel tracing framework to connectfractured vessel segments. Experiments on four CT image cases are conductedand analyses are also carried out. Substantial improvements are shown bycomparing this vessel extraction method to the method based on vascular activecontour and to the method based on the Hessian matrix. |
PDF Full Text Request