Automatic multi-atlas segmentation using dual registrations

Atlas-based segmentation is widely used in medical image analysis. In multi-atlas segmentation, segmentations of multiple atlas images are propagated to a target image through image registration and the results are combined to obtain a final segmentation for the target image. However, registration does not always perfectly work between the target and each atlas image. Furthermore, the preparation of large number of atlases is also not easy. In this paper, we utilized two different characterized registration methods for deforming each atlas in multi-atlas segmentation. Since deformed atlases are voxel-wise candidates for a final segmentation mask, the proposed method has a similar effect to having more atlas sets, from which we could improve segmentation accuracy and reliability.