CSAIL Publications and Digital Archive header
bullet Research Abstracts Home bullet CSAIL Digital Archive bullet Research Activities bullet CSAIL Home bullet

link to publications.csail.mit.edu link to www.csail.mit.edu horizontal line


Research Abstracts - 2007
horizontal line

horizontal line

vertical line
vertical line

Joint Registration and Segmentation with Multiple Atlases

Boon Thye Thomas Yeo, Mert Sabuncu & Polina Golland


We are developing a Bayesian framework for the use of multiple atlases in the joint registration and segmentation of medical images


Registration is a fundamental problem in medical imaging. The purpose of registration is to bring images into alignment to allow for meaningful comparisons. One of the biggest challenges in medical imaging is to align multiple images, typically from multiple individuals together. In particular, multi-image registration usually assumes the existence of an atlas (either implicitly or explicitly), which is simply a model that summarizes the captured information about the population. The idea behind the use of multiple atlases is the observation that a single atlas might be insufficient to represent the entire population. Previous literature suggests the possible existence of subpopulation modes in normal human brain MRI images [1]. However, it is unclear from how one could exploit these modes. In particular, given a new brain image, a yet to be answered question is: which mode does the new brain image belong to?

Our Bayesian formulation extends the joint registration-segmentation framework of Poh et al. [2] to include the use of multiple atlases. While our framework is general, the particular problem we will test is the parcellation of cortical surfaces. Unlike the subpopulation atlases hinted in previous work [1], we will consider multi-scale atlases obtained from training images registered under various degrees of rigidity (these multi-scale atlases are so named because they become increasingly sharper as more flexible warps are employed). In canonical non-linear image registration, one optimizes an objective function that consists of an image fidelity term and a regularization term. However, the trade-off between the two terms is set arbitrarily. In particular, if we favor regularization, then the images are warped with a constrained family of transformations (i.e. “rigid-like”). On the other hand, favoring the image fidelity term will result in the images warping substantially to match the atlas. One of the aims of this work is to automatically select the trade-off between the regularization and image fidelity.

Another reason for the use of multi-scale atlases is the theoretical consistency between the atlases and the test images. When we apply a sharp atlas (obtained from training images that have been registered under minimally constrained warps) to a new test image, an underlying assumption is that the new image is sampled from the generative model of the atlas [3]. However, since the sharp atlas is trained from images which are already registered, this assumption is clearly not true. These increasingly sharp atlases therefore act as waypoints to guide the test image towards the final warps, and thus increase the capture range of the registration algorithm.

Preliminary Results

Parcellation examples on a partially inflated surface. Brown lines are the ground truth boundaries. Blue lines are the automated segmentation boundaries.


Bruce Fischl (MGH)


Boon Thye Thomas Yeo is funded by the Agency for Science, Technology and Research, Singapore. Support for this research was provided in part by NIH NIBIB NAMIC U54-EB005149, NIH NCRR NAC P41-RR13218, NIH NINDS R01-NS051826, NIH NCRR mBIRN U24-RR021382, P41-RR14075, R01 RR16594-01A1 and NSF JHU ERC CISST, NCRR P41-RR14075, R01 RR16594-01A1, the National Institute for Biomedical Imaging and Bioengineering (R01 EB001550), the National Institute for Neurological Disorders and Stroke (R01 NS052585-01) as well as the Mental Illness and Neuroscience Discovery (MIND) Institute.


[1] Daniel J. Blezek, James V. Miller: Atlas Stratification. MICCAI (1) 2006: 712-719

[2] K. M. Pohl, J. Fisher, W.E.L. Grimson, R. Kikinis, and W.M. Wells. A bayesian model for joint segmentation and registration. NeuroImage, 31(1), pp. 228-239, 2006

[3] L . Zöllei: "A Unified Information Theoretic Framework for Pair- and Group-wise Registration of Medical Images", Ph.D. thesis, MIT


vertical line
vertical line
horizontal line

MIT logo Computer Science and Artificial Intelligence Laboratory (CSAIL)
The Stata Center, Building 32 - 32 Vassar Street - Cambridge, MA 02139 - USA
tel:+1-617-253-0073 - publications@csail.mit.edu