Improved Sparse Shape Composition Model for Multi- shape Prior

Shape prior modeling is a challenging and crucial component in various image segmentation applications. Most existing methods aim at dealing with single object’s shape variation, which are not directly applicable for multishape prior modeling. In this paper, we present an extension of recently proposed Spare Shape Composition model (SSC) for multi-shape prior modeling. In this extension, multiple shapes of one patient are regarded as a group. A sparse linear composition of training groups is computed iteratively to infer/refine the input group. Thus, not only the a-priori information of each shape but also the a-priori codependency information among different shapes is implicitly incorporated on-the-fly. To validate the efficacy of our method, a 2D left ventricular endocardium and epicardium localization experiment was conducted. The localization result demonstrates that the utilization of our method can achieve more accurate and stable localization compared with SSC.