Anatomical landmarks can play key roles in medical image understanding including segmentation. For example, statistical shape model-based segmentation can be enhanced with landmark information which helps parameter initialization such as pose and locations of models. We have been working on local appearance-based landmark detection scheme. When we define landmarks with surrounding appearance in medical images, certain uncertainty is observed depending on local intensity structures around the landmark. It is obvious that good landmarks should have low uncertainty and also that uncertainty causes difficulty in consistent evaluation of landmark localization error. In this paper, we describe our method for landmark uncertainty quantification based on arrival times of level-set evolution named appearance similarity flow, controlled by similarity between landmark appearance and that of the location within whole image. By using 12 clinical CT dataset, the method was evaluated.
展开▼
