Label fusion using performance estimation with iterative label selection

Atlas-based segmentation is a well-known method of automatically computing segmentation. When multiple atlases are available, then each atlas can be used to compute a 'label', which is an estimation of the ground truth segmentation of a target image. By combining these labels, a more accurate approximation of the ground truth segmentation can be made. A common method to combine labels is the STAPLE algorithm, but this method fails when the performance of the labels highly varies. Other methods select labels based on their estimated performance, but combine them using a simple majority vote procedure. In this paper, a simpler variant of the STAPLE algorithm is presented that iteratively selects labels. Results are given that show that the proposed method outperforms STAPLE in an application to segmentation of the prostate.