Active Contour Models And Application In Virtual Endoscopys

Image segmentation is the foundation of Virtual Endoscopy (VE), and path planning is a prerequisite task to automatically navigate with virtual camera. Both of them play a crucial role in virtual endoscopy application. The dissertation focuses on studying the two key techniques of VE.The objects diagnosed in VE are those tubular structures which are always thin and long. Traditional segmentation methods are not efficient to segment those tubular objects; the segmentation results are always coarse or hollow. Although the Active Contour Model (ACM) could get a smooth and continuous object contour, there are leakage boundaries when extracting surfaces of those long-thin structures. Furthermore the existing path planning methods are high computational cost, which hold back the using of VE. So image segmentation and path planning methods are systematically analyzed and studied based on the active contour model.Firstly, for the weak edges leakage problem existing in the VFC Snake model, an improved VFC Snake model is proposed. The magnitude of the vector field kernel is redefined, which reduces the sensitivity between the capture range of the external forces and the change of parameter. And that a dynamic force, the combination of the VFC force field with the potential force field, is constructed in order to alleviate the weak edges leakage problem. Experiments demonstrate the improved VFC Snake model is robust to noise and can converge to the object boundary accurately.Due to parametric ACM can not handle topological change, the VFC fields are generalized to the geometric ACM which has topology-adaptable ability. A VFC geometric active contour model is presented. The model minimizes the model energy by two phases. In the first phase, the LBF method is used to optimize the model based on the image region information, which could get the image global optimum information. In the second phase, VFC force field pulls the model contours to the object boundary with concave regions when the evolvement speed of the model is slow. Since the long-range capture, VFC force field can pull the contours lying in or out the object boundary to the object boundary, the model can avoid the weak edges leakage problem which is an annoyance in traditional Level Set methods. Experimental results show the model is effective.Secondly, when the surfaces of the tubular structures are extracted by the minimal action surface, edges leakage problem could not avoided. So a front propagation freezing criterion is developed based on the average energy of the front propagation. To freeze the"tail"of the front and its energy less than the average energy of the whole front propagation, and to continue propagate the"head"of the front and its energy larger than the average energy of the whole front propagation. While the"head"front gets the end of the objects, the whole surfaces of the object are extracted. The criterion is based on the global optimum, so can enhance the robustness to the noise and low contrast medical images. Experimental results show the criterion avoids the edges leakage problem effectively.Thirdly, a new centerline planning algorithm based on B-Snake model is designed. The centralized method of moving regular polyhedron is defined, which provides a centralized force as the exterior force of B-Snake model. And the interior force of the B-Snake is omitted, so it is ease to control the model. With the characteristic of B-spline, a few control points are chosen which could get the smoothing and continuous centerlines. So the computer time is reduced largely. Another advantage of the algorithm is that it is not limited by the segmentation results and could be executed in original images data directly.Finally, a path planning algorithm based on segmentation is built. The algorithm makes use of the connectivity-preserving features of sequence images and segments a series of images by applying the adaptive region growing algorithm. It stores the seed of growing region into a series of stack as the key point of the navigation plan. The key point of navigation path is selected during segmentation, and the navigation path is obtained after smoothing those points. Experiment results show that the algorithm is feasible and robust.