Learning of Atlas Forest Hierarchy for Automatic Labeling of MR Brain Images

Abstract

We propose a multi-atlas-based framework to label brain anatomies in magnetic resonance (MR) images, by constructing a hierarchical structure of atlas forests. We start by training the atlas forests in accordance to individual atlases, and then cluster atlas forests with similar labeling performances into several groups. For each group, a new representative forest is re-trained, based on all atlas images associated with the atlas forests in the group, as well as the tentative label maps output by the clustered atlas forests. This clustering and re-training procedure is conducted iteratively to obtain a hierarchical structure of atlas forests. When applied to an unlabeled image for testing, only the suitable trained atlas forests will be selected from the hierarchical structure. Hence the labeling result of the test image is fused from the outputs of selected atlas forests. Experimental results show that the proposed framework can significantly improve the labeling performance compared to the state-of-the-art method.