30.12.2014 Views

Automatic Registration-Based Segmentation for ... - NeoBrainS12

Automatic Registration-Based Segmentation for ... - NeoBrainS12

Automatic Registration-Based Segmentation for ... - NeoBrainS12

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Automatic</strong> <strong>Registration</strong>-<strong>Based</strong> <strong>Segmentation</strong> <strong>for</strong><br />

Neonatal Brains Using ANTs and Atropos<br />

Jue Wu and Brian Avants<br />

Penn Image Computing and Science Lab,<br />

University of Pennsylvania, Philadelphia, USA<br />

Abstract. We implemented an automatic open-source processing pipeline<br />

<strong>for</strong> neonatal brain tissue segmentation. The framework makes use of N4<br />

to correct bias field, the de<strong>for</strong>mable registration SyN to warp a public<br />

neonatal template to the test image, and multivariate MRF-based segmentation<br />

Atropos to per<strong>for</strong>m segmentation. The pipeline does not need<br />

training and runs efficiently. It can achieve satisfactory results <strong>for</strong> most<br />

classes of a neonatal brain MR image.<br />

1 A Prior-<strong>Based</strong> <strong>Segmentation</strong> Algorithm<br />

We contribute a template-based method <strong>for</strong> the challenging problem of tissue<br />

segmentation in the neonatal brain. The task is difficult because of intensity<br />

inversion of unmyelinated white matter and low resolution (or large slice thickness)<br />

in standard neonatal T1 or T2-weighted MRI. Our framework employs SyN<br />

diffeomorphic registration and Atropos prior-based multivariate segmentation<br />

methods available within the open-source, multi-plat<strong>for</strong>m Advanced Normalization<br />

Tools (ANTs). SyN and Atropos are general-purpose tools, which were not<br />

designed specifically <strong>for</strong> neonatal brain processing. However, with the assistance<br />

of a neonatal brain template, the general purpose methods are adapted <strong>for</strong> a<br />

task-specific algorithm.<br />

There are four components in the proposed framework: registration, bias<br />

correction, segmentation and post-processing (see Fig. 1).<br />

1.1 <strong>Registration</strong><br />

Assuming there is a proper template available <strong>for</strong> each test image to be segmented,<br />

we aim to align the template with the test image and thus trans<strong>for</strong>m<br />

the priors associated with the template to the subject space. The registration<br />

is per<strong>for</strong>med via the diffeomorphic mapping methods available within ANTs<br />

software [1], which rank among top per<strong>for</strong>mers in terms of registration accuracy<br />

[2]. ANTs is freely available at http://picsl.upenn.edu/ANTS/download.php. We<br />

use the neighborhood cross correlation as similarity measure <strong>for</strong> the mono-modal<br />

registration and mutual in<strong>for</strong>mation <strong>for</strong> intra-subject affine registration between<br />

T1 and T2 weighted MR images. <strong>Registration</strong> between T1 and T2 weighted<br />

modalities enables us to employ a multivariate data-term in the segmentation


2 Jue Wu and Brian Avants<br />

Fig. 1. Overview of the proposed algorithm<br />

<strong>for</strong>mulation. The brain mask from the template was trans<strong>for</strong>med to subject space<br />

and helped remove non-brain tissues.<br />

The line of code <strong>for</strong> registration between template and test image is:<br />

ANTS 3 -m CC[t2template-GA.nii.gz, test_image.nii.gz, 1, 2]<br />

-r Gauss[3,0] -t SyN[0.25] -i 90x100x80 -o output.nii.gz<br />

1.2 Template Used<br />

To a large extent, the success of prior-based segmentation hinges on the quality<br />

of the chosen template and the similarity between template and test image. We<br />

chose a set of premature neonatal T2 templates from Imperial College London,<br />

which have high definition and good tissue contrast with various post-menstrual<br />

ages. These atlases are based on 204 preterm babies with no observed obvious<br />

pathology and freely available at http://www.brain-development.org/. We<br />

picked two templates, one 30 and the other 40 week old, <strong>for</strong> the current processing.<br />

The original templates have no label <strong>for</strong> myelinated WM and no separated<br />

labels <strong>for</strong> ventricular CSF and extra-cerebral CSF so we created the labels manually<br />

according to the Challenge’s protocol.<br />

1.3 Pre-processing<br />

Pre-processing be<strong>for</strong>e segmentation only involved MR field inhomogeneity correction.<br />

We employed the N4 method [3] to correct bias field and used the probabilistic<br />

map of white matter as a reference. Bias field correction was done <strong>for</strong><br />

independently <strong>for</strong> both T1 and T2 images.


Title Suppressed Due to Excessive Length 3<br />

1.4 <strong>Segmentation</strong><br />

Tissue classification is accomplished by Atropos which uses expectation maximization<br />

to optimize a probabilistic <strong>for</strong>mulation that links a data term with both<br />

spatial and Markov Random Field (MRF) priors. The posterior probability in<br />

the MRF model is composed of the likelihood probability <strong>for</strong> voxel intensity,<br />

prior probability <strong>for</strong> smoothness constraint and template priors. Optimization<br />

is achieved by expectation maximization with an iterated conditional modes parameter<br />

update schema. For technical details about Atropos, readers are referred<br />

to the work of Avants et al. [3]<br />

The Atropos method takes prior probability maps of all tissue classes and<br />

the T2 image (plus the T1 <strong>for</strong> the final iteration) as input. It outputs the segmentation<br />

of all classes as well as probability maps <strong>for</strong> all classes. We restarted<br />

this process three times by setting the output probability maps of the previous<br />

iteration as input probability maps of the next iteration. Due to close intensity<br />

profiles between several tissues (such as cortex and deep gray matter, white matter<br />

and cerebellum, ventricular CSF and extra-cerebral CSF), the tissue priors<br />

were recombined with output probability maps between two iterations to rein<strong>for</strong>ce<br />

spatial constraint. Further details are available in the scripts used <strong>for</strong> this<br />

study which are available from the authors.<br />

The line of code <strong>for</strong> one iteration of prior-based segmentation:<br />

Atropos -d 3 -x mask.nii -c [2,1.e-9] -m [0.10,1x1x1] -u 0 --icm [1,1]<br />

-a test_t1.nii -a test_t2.nii -o [seg.nii.gz,segProb%02d.nii.gz]<br />

-i PriorProbabilityImages[8,prior%02d.nii.gz,0.5] -p Socrates[0]<br />

1.5 Post-processing<br />

The purpose of post-processing is to decrease the number of misclassified voxels<br />

due to unusual intensity of these voxels. The MRF smoothness constraint may<br />

not be able to remove them because they can appear in a cluster. These clusters<br />

can be reduced by reimposing the template-based priors, which are likely to show<br />

low probability of the tissue at this location. There<strong>for</strong>e the misclassification was<br />

replaced with higher probability tissue class.<br />

1.6 Strengths and Limitations<br />

The proposed algorithm has several advantages.<br />

No training is needed. The algorithm only needs a labeled template, where<br />

prior knowledge of tissue distribution is incorporated, and thus training examples<br />

are not necessary.<br />

Relatively fast to run. For each subject, the registration unit takes 40 to 60<br />

minutes depending on the size of the brain. The segmentation part takes about<br />

20 minutes. Bias correction is about 20 minutes and post-processing lasts lest<br />

than 1 minute. The computation was done on a Linux plat<strong>for</strong>m with Intel Xeon<br />

3.2GHz CPU and 2GB memory.


4 Jue Wu and Brian Avants<br />

Open source and public templates. The implementation is based on opensource<br />

softwares and we used templates freely available to the public. There<strong>for</strong>e<br />

the proposed method should have good reproducibility and can be made available<br />

to the public in the future.<br />

One significant limitation of the proposed method is that the extra-cerebral<br />

CSF segmentation is suboptimal due to the fact that the template did not have<br />

features outside of the cerebrum (skull, dura, etc) whereas the test images cover<br />

the whole head. While our registration methodology is fairly robust, in some<br />

cases, this difference may be significant enough to impair registration per<strong>for</strong>mance.<br />

In such cases, the segmentation will also suffer. It is also possible that<br />

neither smoothness prior and template prior can eliminate all misclassification.<br />

Additional anatomical constraints may be helpful, <strong>for</strong> example, WM and extracerebral<br />

CSF voxels cannot be neighbors.<br />

2 Evaluation Results<br />

The snapshots <strong>for</strong> some example segmentation are shown in Fig. 2. The quantitative<br />

evaluation results from the Challenge are shown in Fig. 3, 4 and 5.<br />

Fig. 2. First row is one slice of the first subject from each set (40wk axial, 30wk coronal,<br />

40 wk coronal. Second row is the corresponding segmentation.


Title Suppressed Due to Excessive Length 5<br />

Fig. 3. Quantitative results <strong>for</strong> set1


6 Jue Wu and Brian Avants<br />

Fig. 4. Quantitative results <strong>for</strong> set2


Title Suppressed Due to Excessive Length 7<br />

Fig. 5. Quantitative results <strong>for</strong> set3


8 Jue Wu and Brian Avants<br />

References<br />

1. Avants, B., Epstein, C., Grossman, M., Gee, J.: Symmetric diffeomorphic image<br />

registration with cross-correlation: Evaluating automated labeling of elderly and<br />

neurodegenerative brain. Medical Image Analysis 12 (2008) 26–41<br />

2. Klein, A., Andersson, J., Ardekani, B., Ashburner, J., Avants, B., Chiang, M., Christensen,<br />

G., Collins, D., Gee, J., Hellier, P., Hyun, S., Jenkinson, M., Lepage, C.,<br />

Rueckert, D., Thompson, P., Vercauteren, T., Woods, R., Mann, J., Parsey, R.:<br />

Evaluation of 14 nonlinear de<strong>for</strong>mation algorithms applied to human brain mri registration.<br />

NeuroImage 46 (2009) 786–802<br />

3. Avants, B., Tustison, N., Wu, J., Cook, P., Gee, J.: An open source multivariate<br />

framework <strong>for</strong> n-tissue segmentation evaluation on public data. Neuroin<strong>for</strong>matics 9<br />

(2011) 381–400

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!