Atlas Forests Multi-Atlas Label Propagation with Atlas Encoding by Randomized Forests

We describe our submission to the MICCAI 2013 SATA Challenge. The method is based on multi-atlas based label propagation, its major characteristic being that is uses the concept of an atlas forest to represent an atlas. This results in an efficient scheme, which requires only a single registration to label a target. Fusion of the probabilistic label proposals from each atlas is done by averaging across atlases. Results are submitted for the unregistered Diencephalon data set. 1 Method Overview This submission is based on the concept of atlas forests, which is presented in detail in [1]. Therefore, we describe the approach only at a high-level, and list specific modifications of [1]. The submitted approach follows the standard multi-atlas label propagation model. For a given target image, a label proposal is generated by each atlas, and the proposals are then fused into a final labeling. The special characteristic of our approach is that each atlas is represented by a randomized classification forest, which is trained only on this atlas. We call this an atlas forest (AF). The approach is designed with the goal of efficiency. To label a target, only a single registration is needed, which is in contrast to most existing approaches, which require the registration of all atlases to the target. This registration aligns a probabilistic atlas to the target, and the aligned label priors are then used to augment the original input for the processing by the AFs. The actual evaluation of the atlas forests is also highly efficient. At test time, each atlas forest produces a probabilistic label estimate, which are then fused by averaging. The approach is summarized in Fig. 1. Compared to [1], we modified the system for the challenge submission in two points, due to the comparably small number of labels. This allowed us to use context-sensitive features for label priors, now effectively treating prior and intensity channels in the same way ([1] used only local features were used on priors). Also, we did not use hierarchical priors for the challenge submission. Finally, we tested two new variations. The first artificially augments the training set by left/right mirroring of each atlas. In the challenge system, submissions marked Attempt No. 1/2 are based on the original/augmented training set. 2 Darko Zikic, Ben Glocker, and Antonio Criminisi Training: Encoding each atlas by an Atlas Forest atlas forest 1 atlas forest 2 atlas forest 3 atlas forest n ... ... inner/outer left/right