Here, we propose a confidence shape metric for level set segmentation. First, the confidence shape metric, which encodes local confidence into the matching result, is used in matching shapes and producing confidence maps. Then, based on the confidence shape prior, the level set function evolves to a desired contour. The proposed shape metric allows representation of shape variations beyond the coverage of training shapes under the level set framework, which is suitable for segmenting strongly deformed and cluttered images, especially when the set of training shapes is sparse compared with numerous intracategory variations. We evaluated the proposed approach on the challenging Weizmann dataset and computed tomography images. Experimental results indicate the advantage of confidence shape prior over shape prior without confidence under the Dice-coefficient metric.