Classifier-Based Multi-atlas Label Propagation with Test-Specific Atlas Weighting for Correspondence-Free Scenarios

We propose a segmentation method which transfers the advantages of multi-atlas label propagation (MALP) to correspondence-free scenarios. MALP is a branch of segmentation approaches with attractive properties, which is currently applicable only in correspondence-based regimes such as brain labeling, which assume correspondence between atlases and test image. This precludes its use for the large class of tasks without this property, such as tumor segmentation. In this work, we propose a method which circumvents the correspondence assumption by using a classifier-based atlas representation in the spirit of the recently proposed Atlas Forests (AF). To counteract the negative effects of the over-training property of AF for applications with highly heterogeneous examples, we employ test-specific atlas weighting by the STAPLE approach. The main idea is that over-training ceases to be a problem if the prediction is based only on training atlases which are “similar” to the test image. Here, the “similarity” is based on the estimated ability of an atlas-based classifier to perform a correct labeling. We show a successful use of the proposed method for segmentation of brain tumors on data from the BraTS 2013 Challenge, which presents a correspondence-free scenario in which standard MALP cannot be expected to operate.