Interactive Registration and Segmentation for Multi-Atlas-Based Labeling of Brain MR Image

Abstract

In the conventional multi-atlas-based labeling methods, atlases are registered with each unlabeled image, which is then segmented by fusing the labels of all registered atlases. The registration is typically ignorant about the segmentation while the segmentation of each individual unlabeled image is independently considered, both of which potentially undermine the accuracy in labeling. In this work, we propose the interactive registration-segmentation scheme for multi-atlas-based labeling of brain MR images. First, we learn the distribution of all images (including atlases and unlabeled images) and register them to their common space in the groupwise manner. Then, we segment all unlabeled images simultaneously, by fusing the labels of the registered atlases in the common space as well as the tentative segmentation of the unlabeled images. Next, the (tentative) labeling feeds back to refine the registration, thus all images are more accurately aligned within the common space. The improved registration further boosts the accuracy to determine the segmentation of the unlabeled images. According to our experimental results, the iterative optimization to the interactive registration-segmentation scheme can improve the performances of the multi-atlas-based labeling significantly.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.