Confident prior guided level-set segmentation using adaptive external force

In this paper we propose a new level-set based image segmentation method. Our contributions lie in two aspects. Firstly, to solve initialization sensitivity of existing methods, we propose an adaptive external force. It can dynamically determine its direction according to the change of contrast between two sides of the curve during the evolving process, in contrast to balloon force in traditional snakes which either contract or expand according to initial condition. Secondly, we propose a shape matching criterion and use it to select the mostly matched template and evaluate its confidence value. We integrate these two techniques into level-set based segmentation. The resulting contours are more accurate due to the fact that large intra-class deformation and clutter has been considered in confident shape prior. We compare experimental results with other deformable models and two state-of-art models. We analyze advantages and disadvantages of our model.