Accurate segmentation for right ventricles based on affinity propagation clustering and multi-Atlas selection

Due to the unique characteristics of right ventricle, such as volatile, thin wall, unobvious boundary, multi-Atlas algorithm is appropriate for its segmentation. However, most of the existing Atlas select methods are based on choosing Atlas after registering, while the registering is time consuming and reduce the segmentation performance. For this reasons, we introduce a new Multi-Atlas selection method based on affinity propagation clustering algorithm. Firstly, see all Atlas images as a series of data points, clustering them through message propagation. Secondly, register all the clustering centre images to target image, getting deformation markers results by STAPLE label fusion. Finally, sort all the fusion results by dice similarity coefficient. Register and fusion the clustering center images which own the biggest dice similarity coefficient. Furthermore, repeating the process above until get accurate segmentation. Experiment results show that the proposed method can segment right ventricle effectively. Compared to the traditional selection ways, segmentation accuracy has been greatly improved through this method.