Abstract: Multialtas segmentation methods are currently popular methods to tackle the unique characteristics of right ventricles， e.g. volatile， thin and unobvious boundary. Most existing methods select the best atlas after the whole registration， which makes the process time consuming and poor performing. A new method was proposed based on affinity propagation clustering algorithm. First， all atlas images were treated as a series of data points and were clustered through message propagation. Further， all the exemplars were registered to the target image， and the label image was deformed by the parameters obtained from the registration process. What’s more， the label image was fused by STAPLE label fusion method. At last， the exemplars were sorted according to dice coefficient similarity and were used with largest similarity in the registration and segmentation. The process was repeated until reaching an accurate segmentation. Experimental results showed that the proposed method could segment right ventricle effectively. Comparing to the traditional selection methods， the segmentation accuracy has been considerably improved.

Key words: right ventricle, image segmentation, atlas, affinity propagation clustering, image registration