2630
A knowledge-based linear registration for brain MRI morphology1Guangdong Provincial Key Laboratory of Medical Image Processing, School of Biomedical Engineering, Southern Medical University, Guangzhou, China, 2Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, United States, 3F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, MD, United States
Synopsis
Linear registration is an essential first step for image registration. However, linear registration often fails when the brain shapes, locations, orientations of the target and template images are severely different. To solve this problem, we proposed a knowledge-based approach, in which a large number of MR images were prepared as intermediate images, which were semi-automatically registered to the template a priori to ensure accurate registration. A new target image was first registered to all intermediate images and best intermediate image was selected based on a goodness-of-fit metric. The final transformation was then calculated by combining the pre-determined intermediate-to-target transformation.
Introduction
Registration is widely used in the field of medical image analysis[1]. It is common to first perform rough alignment using linear registration, followed by accurate non-linear registration. Thus, the quality of initial linear registration is crucial. Unfortunately, linear registration easily fails when the appearances of two images are high different. To avoid this problem, much efforts have been made to develop sophisticated registration algorithms [2, 3]. Even so, we expect a certain rate of failure. In this work, we proposed a complementary approach to improve linear registration accuracy, which is based on knowledge databases.Methods
Template: The JHU-MNI-SS T1 image was used as the target [4]. The voxel size was 181x217x181 and the resolution was 1x1x1 mm3.
Test data and Mediator: The test and intermediate were obtained from JHU T1 atlas library. These images are previously segmented and we used segmentation masks to quantitatively judge the quality of registration. 144 subject T1-weighted images were randomly selected and divided into two groups: 96 images as test images and 48 as mediators.
The proposed knowledge-based method is illustrated as in Fig.1. First, we prepared N T1 images (48 mediators) with different brain appearances. The transformation from each mediator to the atlas were solved in advance. Then, the affine transformation of AIR[5, 6] was performed to register the test images to each mediator. The sum of squared difference (SSD) was adopted as a measure to select the best mediator for each subject image. Then, the composition of the linear and per-determined transformation performed to transform the subject image and the associated labels into the template space.
We use DICE metrics to measure the registration quality. We also tested if the number of mediators could be reduced without sacrificing the registration accuracy.
Results
Fig. 2 shows the Dice values for pairwise registration, registration with a single mediator and best mediator selected by SSD for the 96 test images. The mean Dice and success rate by using best mediator is significantly increased, compared to the other two linear registration methods. However, this method was not perfect. For example, for subject #21 (indicated by a star), the proposed method had a result (Dice=0.8) inferior to a single mediator (Dice=0.87).
Table 1. shows the performance of linear registration with automatically selected mediator by SSD from complete 48 mediators and reduced mediator sets. The registration accuracy is slightly decreased from complete 48 atlas to 25 clusters. The further reduction led to clear deterioration of the performance. This result suggests that the method could be twice as fast without severe penalty in the accuracy and, thus, the reduction strategy was effective.
Discussion and Conclusion
A linear registration based on knowledge dataset is proposed. The mediator selection strategy improved the linear registration performance. Please note that we used a simple single-step registration using AIR for this analysis. The proposed approach could be readily combined with more sophisticated multi-step tools to further reduce the occasional failures.Acknowledgements
This publication was made possible by grant R01NS084957 from the National Institutes of Health (NIH). The contents of this paper are solely the responsibility of the authors and do not necessarily represent the official view of NIH or the authors’ affiliated institutions. We thank all the families for their willingness to participate in this study. The authors declare no conflicts of interest.References
1. Zitova, B. and J. Flusser, Image registration methods: a survey. Image and Vision Computing, 2003. 21(11): p. 977-1000.
2. Shekhovtsov, A., I. Kovtun, and V. Hlavac, Efficient MRF deformation model for non-rigid image matching. Computer Vision and Image Understanding, 2008. 112(1): p. 91-99.
3. Myronenko, A. and X.B. Song, Intensity-Based Image Registration by Minimizing Residual Complexity. Ieee Transactions on Medical Imaging, 2010. 29(11): p. 1882-1891.
4. Oishi, K., et al., Atlas-based whole brain white matter analysis using large deformation diffeomorphic metric mapping: application to normal elderly and Alzheimer's disease participants. NeuroImage, 2009. 46(2): p. 486-99.
5. Woods, R.P., et al., Automated image registration: I. General methods and intrasubject, intramodality validation. Journal of computer assisted tomography, 1998. 22(1): p. 139-52.
6. Woods, R.P., et al., Automated image registration: II. Intersubject validation of linear and nonlinear models. Journal of computer assisted tomography, 1998. 22(1): p. 153-65.
Figures
