A hierarchical local region-based sparse shape composition for liver segmentation in CT scans

Motivated by the goals of improving segmentation of challenging liver cases containing low contrast with neighboring organs and presence of pathologies as well as highly varied shapes between subjects, a novel framework is presented for liver segmentation in portal phase of abdominal CT images. In a first training step, we describe a multilevel local region-based Sparse Shape Composition (SSC) model, called MLR-SSC, to increase the flexibility of shape prior models and capture the detailed local shape information more faithfully. Specifically, the liver shapes are decomposed into multiple regions in a multilevel fashion. Moreover, we build a local shape repository for each region and refine an input shape in a region-by-region manner. In a second testing step, it starts with a blood vessel-based liver shape initialization to derive a more patient-specific initial shape, followed by a hierarchical deformable shape optimization algorithm. It makes the segmentation framework more efficient and robust to local minima. Extensive experiments on 60 clinical CT scans demonstrate that our method achieves much better accuracy and efficiency than two closely related methods in the presence of small training sets. Moreover, our method shows slightly superior performance to three newly published methods. Also, we compare our method with the published semi-automatic methods from the "MICCAI 2007 Grand Challenge" workshop. HighlightsWe propose a multilevel local region-based Sparse Shape Composition shape model.We present a blood vessel-based liver shape initialization method.We employ a hierarchical optimization strategy to make the framework efficient.The framework is successfully applied to segment liver tissue from CT images.