Dynamic Tree-Based Large-Deformation Image Registration for Multi-atlas Segmentation
Multi-atlas segmentation is a powerful approach to automated anatomy delineation via fusing label information from a set of spatially normalized atlases. For simplicity, many existing methods perform pairwise image registration, leading to inaccurate segm
- PDF / 673,701 Bytes
- 9 Pages / 439.37 x 666.142 pts Page_size
- 9 Downloads / 181 Views
Abstract. Multi-atlas segmentation is a powerful approach to automated anatomy delineation via fusing label information from a set of spatially normalized atlases. For simplicity, many existing methods perform pairwise image registration, leading to inaccurate segmentation especially when shape variation is large. In this paper, we propose a dynamic tree-based strategy for effective large-deformation registration and multiatlas segmentation. To deal with local minima caused by large shape variation, coarse estimates of deformations are first obtained via alignment of automatically localized landmark points. A dynamic tree capturing the structural relationships between images is then used to further reduce misalignment errors. Validation on two real human brain datasets, ADNI and LPBA40, shows that our method significantly improves registration and segmentation accuracy.
1
Introduction
Multi-atlas segmentation is an automated approach to delineating anatomical structures of a target image by borrowing complementary information from multiple pre-annotated atlases. Segmentation labels from multiple atlases that are registered to the target image are combined to obtain the ultimate segmentation result. This approach avoids not only time-consuming manual annotation but also potential bias introduced by segmentation with only one single atlas. Due to its promising results, it has been widely used in medical image analysis to segment the brain [1], the heart [2] and abdominal organs [11]. Although much effort has been put to improve the accuracy of label fusion, much less emphasis has been put on image registration. Most of the existing multi-atlas segmentation methods merely perform simple pairwise registration [1,2] by aligning each atlas independently to the target image. This approach fails to consider the correlation between atlases and thus leads to inconsistency among the atlases when labeling the same anatomical structure. Simple pairwise D. Shen—This work was supported in part by a UNC BRIC-Radiology start-up fund, and NIH grants (EB006733, EB008374, EB009634, MH088520 and NIHM 5R01MH091645-02). c Springer International Publishing Switzerland 2016 B. Menze et al. (Eds.): MCV Workshop 2015, LNCS 9601, pp. 137–145, 2016. DOI: 10.1007/978-3-319-42016-5 13
138
P. Zhang et al.
registration also does not take advantage of the structural similarity between images to help overcome registration related problems such as local minima. Such problem occurs quite often when images from different populations (i.e., patients and healthy controls) vary dramatically in anatomical structures. There have been some recent attempts to overcome the above problems by using more sophisticated registration strategies. For example, Hoang Duc et al.[5] attempted to establish the relationship across atlases and the target image by iteratively registering them to an evolving group mean image [8]. However, this only works well when images in the group can be registered reasonably well. Also, the mean image is sensitive to registration outl
Data Loading...
