Segmentation of Brain MRI in Young

Children

Maria Murgasova a Leigh Dyet b David Edwards b

Mary Rutherford b Jo Hajnal b Daniel Rueckert a

aVisual Information Processing Group, Department of Computing, ImperialCollege London

bDepartment of Imaging Sciences, Faculty of Medicine, Imperial College London

Abstract

This paper deals with an automatic tissue segmentation of brain MRI in youngchildren. We examine the suitability of state-of-the-art methods developed for theadult brain when applied to the segmentation of the brain MRI in young children.We develop a method of creation of a population-specific atlas in young childrenusing a single manual segmentation. The method is based on non-linear propagationof the segmentation into population and subsequent affine alignment into a referencespace and averaging. Using this approach we significantly improve the performanceof the popular EM segmentation algorithm on brain MRI of young children. Themethod can be used for building probabilistic atlases with any number of structures.We compare resulting algorithm with non-rigid registration based label propagation.Finally, the both methods are used to measure the volume of 7 brain structures andmeasure the growth between 1 and 2 years of age.

Key words:PACS:

1 Introduction

The problem of automatic segmentation of brain MRI has been extensivelystudied in past decade thanks to its applications in clinical studies of normaland diseased brain. Recent research of brain development in prematurely bornchildren requires reliable and accurate automatic segmentation techniques forthe neonatal and early childhood brain as well. Approximately 40 % of prema-turely born children have cognitive, neurological or behavioural impairment[1] and the cerebral abnormalities underlying problems such as minor motorimpairment or inattention still remain unclear. Accurate tissue quantification

Preprint submitted to Elsevier 7 March 2012

of the preterm brain in early childhood would allow us to study the growth ofdifferent structures and help to relate developmental outcomes to changes involume and shape of anatomical structures.

In addition to noise, intensity non-uniformity, partial volume effect and nat-ural tissue intensity variation present in the adult brain MRI, neonatal andinfant brain MRI tissue intensities have even greater natural variation causedby on-going process of myelination of white matter. This is most profound inneonates where white matter is extremely difficult to distinguish from greymatter on T1 weighted MRI sequences. By the age of 1 or 2 years, the ma-jority of white matter has myelinated. However, the natural tissue intensitydistribution overlap results in large misclassifications in central brain struc-tures. This problem is further compounded by the fact that child brain differssignificantly in size and shape from the adult brain, in particular in the centralstructures of the brain [2]. This causes problems for methods which rely onprobabilistic atlases built from adults as spatial prior information.

In this paper we focus on the segmentation of brain MRI of 1 and 2-year-oldchildren. We show how to bootstrap a probabilistic tissue atlas for a specificpopulation based on registration of a single manual segmentation to a popula-tion and subsequent alignment to the space of a reference subject and averag-ing. We demonstrate that the results of the expectation-maximization methodby Van Leemput et al. ([3], [4]) on young children can be significantly improvedby using an atlas which appropriately captures the anatomical variability ofthe infant brain and compare the results with non-rigid registration-based la-bel propagation by Rueckert et al. [5]. The both methods are then used tomeasure the volume of 7 brain structures and measure the growth between 1and 2 years of age.

2 Related Work

To date the research in MRI brain segmentation has focused on adult andneonatal brains. The most popular class of tissue classification methods isbased on expectation-maximization framework ([6]; EMS:[3], [4]; SPM: [7],[8]; FAST:[9]) which iteratively interleaves calculation of intensity probabilitydistributions and/or bias field with refinement of soft segmentations of 3 basicbrain tissue classes - white matter (WM), grey mater (GM) and cerebro-spinal fluid (CSF). A mixture Gaussian distribution is assumed as a modelfor the voxel intensity probability distribution as it is easy to incorporate intoEM framework. However, the real distribution can significantly differ fromGaussian, see Fig. 3a. This is caused by partial volume effect as well as naturalintensity variation within the 3 basic tissue classes, see Fig. 1.

2

Natural tissue intensity variation in different areas of the brain is one of thesignificant complications resulting in misclassifications. Fischl et al. [10] pointout that different central brain structures have different intensity distributions,see Fig. 1. GM in central brain structures is characterized by brighter intensi-ties than cortical GM, and is therefore likely to be misclassified as WM. Renet al. [11] reports such misclassification in adult brain using the EM-basedsegmentation method FAST [9]. However non-rigid registration-based labelpropagation [5] succeeds in this area of the brain as it does not assume anymodel for tissue intensities.

Fig. 1. Real tissue distribution of different brain structures based on manual seg-mentation of a 2-year-old child brain MRI. Corpus callosum is a central brain whitematter structure with significantly higher intensities than general white matter.Deep gray matter structures pallidum, thalamus, putamen and caudate are brighterthan cortical gray matter.

Inevitable misclassifications resulting from overlaps in tissue intensity distribu-tions can be partially avoided by spatially constraining the segmentation pro-cess with prior information in form of an aligned probabilistic atlas (Fig. 3a)at each iteration ([3], [4]); however this makes the method very sensitive to thecorrect alignment of the atlas with the image. This is very difficult for childbrains which significantly differ from adult brains in size and shape [2]. Tosolve this problem, [12] and [13] suggest to warp the atlas to subject by non-rigid registration. In [14] Pohl reports difficulties with non-rigidly registeringthe template to image which is consistent with our own experience. Recentmethods try to overcome this problem by iteratively warping a deformableatlas and refining the segmentation at the same time [8], [15].

An alternative non-parametric method for adult brain classification was de-veloped by Cocosco et al. [16]. In this method the atlas is used only to selectthe training samples which are consequently used in kNN classification. Thismethod is reported to deal well with anatomies which differ from the priorinformation. However, it cannot deal with the problem of tissue intensity dis-tribution overlap in single-channel T1 data.

The adult brain segmentation methods can be applied to the neonatal and in-fant brain when tuned to their specific properties. This was shown by Prastawaet al. [17] who used the EM approach ([3], [4]) with an atlas derived from singlesubject segmentation used as a prior, followed by non-parametric classification

3

similar to [16] on neonatal data.

3 Segmentation Methods

3.1 Registration-based segmentation

In registration-based methods a segmented image (deformable atlas) is elas-tically warped to a new image and tissue labels are simply transferred. A de-formable atlas is usually obtained using one or several manual segmentations.The main advantage of this method is ability to segment the brain into anynumber of structures. In addition, intensity variations in the deep GM do notaffect the result as the registration is not guided directly by intensities. How-ever, non-rigid registration algorithms are often not flexible enough to capturecomplex cortical folding or delineate the boundaries of structures precisely. Onthe other hand, invisible structure boundaries can be easily transferred frommanual segmentation.

First, the global 12-parameter affine transformation is calculated to performtranslation, rotation, scaling and skewing so that the best possible alignmentof the images is achieved [18]. Affine registration is not sufficient due to naturalvariability in shape and size of normal healthy brains. The global transforma-tion is therefore followed by local non-linear transformation resulting in elasticor non-rigid transformation. One of the ways to model the local componentis using uniform tensor-product 3D cubic B-splines [5].

(a) (b) (c) (d) (e)

Fig. 2. Registration: (a) the reference image (b) affine registration of the referenceimage to the new image (c) non-rigid registration of the reference image to the newimage (d) the new image (e) non-rigid registration-based segmentation of WM

3.2 EM-based segmentation

The expectation-maximization algorithm (EM) [19] is a general technique forfinding missing data based on observed data and maximum likelihood param-

4

eters estimates. In [4] the observed data are the image intensities, the missingdata are the tissue labels and the parameters are the means and variances ofthe Gaussian distribution which is assumed for the intensity distribution ofeach tissue class. This is an iterative process which interleaves the calculationof soft segmentations for each tissue class (usually white matter, grey matter,cerebrospinal fluid and other; see Fig. 4 c,d,e ) - the expectation step, withmaximum likelihood estimation of the tissue intensity distributions (Fig. 3a)- the maximization step. The segmentation process is constrained by priorinformation in form of a probabilistic atlas (Fig. 3b) which has been alignedwith the image by affine registration.

E-step:

pmik =G(yi, µ

m−1k , σm−1

k )patlasik∑j G(yi, µ

m−1j , σm−1

j )patlasij

M-step:

µmk =

∑i p

mikyi∑

i pmik

σmk =

∑i p

mik(yi − µm

k )2∑i p

mik

where pmik denotes the posterior probability of ith voxel belonging to tissue k atmth iteration, yi is an intensity of ith voxel and G(., ., .) denotes the Gaussiandistribution with the mean µm

k and variance σmk .

(a) (b)

Fig. 3. (a) Intensity distribution of the whole brain, WM, GM and CSF (b) Theprobabilistic atlas created from population of 2-year-old children.

5

(a) (b) (c) (d) (e)

Fig. 4. EM segmentation into 3 tissue classes: (a) the reference image (b) hardsegmentation; soft segmentation: (c) WM (d) GM (e) CSF

4 Prior information for EM segmentation

4.1 Standard probabilistic atlas

As we mentioned above, the reliability of EM-based segmentation dependson correct prior information due to natural intensity variation within WMand GM. In addition to using standard MNI probabilistic atlas, we also usedthe CCHMC pediatric brain atlas created from population of 5 to 9-year-oldchildren [2] (Fig. 9c) with each subject’s image.

We found that when either of those two atlases is aligned with the image,the EM method significantly overestimates WM in the central structures ofthe brain. This is most profound in the thalamus, a deep GM structure, whereabout 30% of voxels were misclassified as white matter (see Fig. 6 and Fig. 9b).The histogram based on the manual segmentation of a 2-year-old 1.0T dataset (Fig. 5) shows that the voxel intensities of this structure lie in the regionof the overlap of white and grey matter intensity distributions and thereforecorrect classification in this area can be achieved only by using correct priorinformation. However, prior information from both probabilistic atlas [2] andstandard MNI atlas overestimates white matter in this area and does notdescribe the anatomy of the brain in early childhood adequately (Fig. 9c).

4.2 Creating a population-specific atlas for EM segmentation

To be able to segment deep GM correctly, we had to create a probabilisticatlas specifically for 1 and 2 year old children. When creating population-specific atlases, obtaining sufficient number of manual segmentations may bevery difficult as the manual delineation of the whole image is extremely time-consuming process. We therefore developed a method for creating a proba-bilistic atlas using only one manual segmentation, see Fig. 7. We obtained amanual segmentation of a single 2-year-old subject to create a probabilistic

6

Fig. 5. Histogram of the tissue intensity probability distribution for WM, GM, CSFand thalamus (a deep GM structure) based on a manual segmentation of brainMRI from a 2-year-old child. Distributions of GM and CSF differ significantly fromthe Gaussian distribution due to partial volume effect. The intensity distribution ofthalamus lies in the region of the overlap of the WM and GM distributions.

(a) (b) (c) (d)

Fig. 6. Misclassification in thalamus: Substantial amount of deep grey matter clas-sified as white. (a) MR image of a 2-year-old brain (b) EM segmentation usingthe adult atlas (c) EM segmentation using the 2-year-old child atlas (d) manualsegmentation

atlas of the 1 and 2-year-old brain. The manual segmentation was then trans-formed to 37 2-year-old subjects by performing non-rigid registration [5]. All37 segmented subjects were then aligned with the reference subject by affineregistration and the segmentations were averaged. The resulting probabilitymaps describe the anatomy of the 2-year-old brain correctly when aligned witha subject by affine registration. We also created a separate probabilistic atlasof the 1-year-old brain. First we non-rigidly warped the manual segmentationof a 2-year-old brain to the scan of the same subject at 1 year to minimize theregistration error. The same process of the atlas creation was then repeatedwith 34 1-year-old scans. The resulting atlases for 1-year-old and 2-year-oldchildren are shown in Fig. 8.

Both of these methods resulted in significant improvement in segmentationof the central brain structures. In addition, this method allows any numberof structures delineated in manual segmentation to be propagated into theresulting probabilistic atlas. (Fig. 6 and 9).

7

Fig. 7. Creating a population-specific atlas. A single manual segmentation is trans-ferred to a population of subjects by non-rigid registration. All the segmentationsare then aligned with the reference subject with affine registration and averaged.

Fig. 8. The comparison of the probabilistic atlases created from population of 1-and 2-year-old children.

5 Comparison of the methods

5.1 Visual inspection

To test the methods we segmented 16 subjects which have been scanned at1 and 2 years of age on 1T HPQ scanner using T1-weighted sequence. We

8

(a) (b) (c) (d) (e) (f)

Fig. 9. Improvement of segmentation using our new atlas: (a) MR image of a2-year-old brain (b) segmentation using the 5 to 9-year-old CCHMC brain atlas(c) the 5 to 9-year-old CCHMC brain atlas (d) our 2-year-old brain atlas (e) seg-mentation using our 2-year-old brain atlas (f) manual segmentation

used both EM-based method with the new population-specific atlas and non-rigid registration-based method. We performed segmentation into standard 3structures (WM, GM, CSF) as well as segmentation into 7 structures (WM,cortex, caudate, putamen, pallidum, thalamus and CSF). The segmentation ofdeep GM is very important task when performing research into effects of thepremature birth on brain development. Before performing the segmentationwe corrected intensity inhomogeneity by the N3 software [20] and removedscalp with Brain Extraction Tool [21]. Also brainstem and cerebellum whereremoved by propagation of labels by non-rigid registration.

Visual inspection of the images confirmed the expectation that EM-basedmethod was more successful in cortical area than registration-based (see Fig. 10b, c). On the other hand the registration-based method has been found morereliable in segmentation of deep GM structures. Those structures exhibitwithin-structure intensity variation and some of the boundaries are not visibleon MRI. Registration-based segmentation does not directly depend on inten-sity and can therefore transfer invisible boundaries from manual segmentation.Also EM-based segmentation of deep GM depends mainly on affine registra-tion of probabilistic atlas and is therefore prone to significant errors. We alsoobserved that that including 7 structures in EM-based segmentation did nothave any significant influence on segmentation of WM and GM compared toonly 3 structures (see Fig. 10 c, d).

It is a known fact that EM-based segmentation with affinely registered prob-abilistic atlas fails on brains with anatomy significantly different from theatlas. We show such failed segmentation on image of a brain severely affectedby focal lesions in Fig. 11c. However, registration-based segmentation can ac-commodate for such a different brain shape (see Fig. 11b).

9

(a) (b) (c) (d)

Fig. 10. Segmentation of brain MRI in 2-year-old child after removing brainstemand cerebellum: (a) the image (b) registration-based segmentation into 7 structures(c) EM-based segmentation into 7 structures (d) EM-based segmentation into 3structures

(a) (b) (c)

Fig. 11. Segmentation of brain MRI in 2-year-old child with focal lesions: (a) the im-age (b) registration-based segmentation into 7 structures after removing brainstemand cerebellum (c) failed EM-based segmentation into 3 structures

5.2 Influence of population-specific atlas on EM-based segmentation into WM,GM and CSF

First, we have validated our results on a single subject for which we obtainedthe manual segmentation. Validation was performed only in the brain area,excluding skull, brainstem and cerebellum. We segmented the image with ourimplementation of the EM method (section 3.2) using 2 different atlases asprior: our new probabilistic atlas based on population of 2-year-old children(2yr atlas) and the standard MNI adult atlas obtained from SPM software [22](adult atlas). For comparison we also segmented the image with EMS software[23] developed by Van Leemput. This software uses the SPM adult atlas, andunlike in our implementation, the bias correction is incorporated within theEM framework. We validated the approach with and without Markov RandomFields (MRF).

The manual segmentation was used as the ground truth. The correctness ofthe segmentation was calculated using Dice metric

D =2 ∗ |Tgt ∩ Tseg||Tgt|+ |Tseg|

(1)

10

where Tgt and Tseg denote the set of samples of given tissue type in the groundtruth and the automatic segmentation, respectively. As we validate the seg-mentation into WM, GM and CSF, the figures for thalamus were obtainedusing following formula:

Dthal =2 ∗ |T thal

gt ∩ TGMseg |

|T thalgt |+ |TGM

seg ∩ T thalgt |

The validation results (Fig. 1a) show that our new atlas improved the segmen-tation of both white and grey matter by 4% compared to EM segmentationwith the adult atlas by both EMS and our implementation. The improvementis most significant in the thalamus where performance increased by 12% incomparison with EMS and by 26% compared to our implementation. Betterperformance of the EMS software in comparison with our implementation ofEM with the adult atlas can be explained by the fact that the bias correctionstep partially compensates for natural intensity variation of within the braintissues.

method WM GM CSF thalamus

EMS 84% 88% 69% 78%

EMS-MRF 85% 89% 62% 80%

EM adult atlas 85% 88% 68% 66%

EM 2yr atlas 88% 92% 78% 92%

Table 1Performance of the segmentation methods on a single 2-year-old subject

We did a similar validation on 6-8 manually segmented slices of additional 4subjects. The average results are given in Table 2. Again, the improvement inWM and GM is only 2%, but it reaches significant 22% in thalamus.

method WM GM thalamus

EM 2yr atlas 86% 88% 92%

EM adult atlas 84% 86% 70%

Table 2Average validation results EM segmentation on four 2-year-old subjects using dice

metrics.

5.3 Comparison of EM-based and registration-based segmentation into 7 struc-tures

We validated EM-based segmentation with the population-specific atlas andregistration-based segmentation on 4 2-year-old subjects with manual segmen-

11

tations of WM, cortex, caudate and thalamus on 5 slices. The results confirmedthe findings of visual inspections. EM-based segmentation is more reliable incortical area and registration-based segmentation performs better in deep GMstructures.

method white cortex caudate thalamus

EM 2yr atlas 87% 87% 84% 86%

NR registration 84% 83% 86% 89%

Table 3Average validation results of EM-based method with population-specific atlas and

non-rigid registration-based label propagation on four 2-year-old subjects using dicemetrics.

6 Calculating brain growth from 1 to 2 years

The methods described above allowed us calculate volumes of 7 structures in16 subjects. As we have scans at both 1 and 2 years of age we could measurethe growth of those structures using two different methods and compare theresults. The average volume measurements for cortex, caudate, putamen, pal-lidum, thalamus and WM are shown in fig. 12. The growth averaged over the16 subjects for each structure is shown in Fig. 13.

Fig. 12. Average volumes of brain structures calculated from 16 subjects at 1 and2 years

Fig. 13. Average growth of brain structures from 1 to 2 years calculated from 16subjects

12

7 Conclusion

In this paper we described the problem of brain MRI segmentation in youngchildren. We analyzed EM-based and registration-based methods and exam-ined their suitability for segmentation of brain MRI during early childhood.We have shown that EM-based method can be successfully used on youngchildren if population-specific atlas is used as a prior. We developed a methodof creation of population-specific atlases from a single manual semgentation.We extended the approach to segmentation of deep GM structures as well.We compared the resulting EM method with registration-based segmentationand calculated volume and growth of brain structures at 1 and 2 years. Wefound that registration-based approach is more reliable in segmentation ofdeep GM structures and can deal with diseased brains, while EM-based ap-proach can deal with complex cortical folding. We also found that the numberof structures did not have any significant influence on resulting EM segmen-tation. Comparison of the growth measurements shows reasonable agreementbetween the 2 methods. However, there is a space for further improvement asmore precise measurements might be required in practice.

References

[1] N. Marlow, D. Wolke, M. A. Bracewell, M. Samara, the EPICure Study Group,Neurologic and developmental disability at six years of age after extremelypreterm birth, Pediatrics 115 (2) (2005) 286–94.

[2] M. Wilke, V. Schmithorst, S. Holland, Normative pediatric brain data for spatialnormalization and segmentation differs from standard adult data, MagneticResonance in Medicine 50 (4) (2003) 749–757.

[3] K. V. Leemput, F. Maes, D. Vandermeulen, P. Suetens, Automated model-based tissue classification of MR images of the brain, IEEE Transactions onMedical Imaging 18 (10) (1999) 897–908.

[4] K. V. Leemput, F. Maes, D. Vandermeulen, P. Suetens, Automated model-based bias field correction of MR images of the brain, IEEE Transactions onMedical Imaging 18 (10) (1999) 885–896.

[5] D. Rueckert, L. I. Sonoda, C. Hayes, D. L. G. Hill, M. O. Leach, D. J. Hawkes,Non-rigid registration using free-form deformations: Application to breast MRimages, IEEE Transactions on Medical Imaging 18 (8) (1999) 712–721.

[6] W. M. Wells III, W. E. L. Grimson, R. Kikinis, F. A. Jolesz, Adaptivesegmentation of MRI data, IEEE Transactions on Medical Imaging 15 (4) (1996)429–442.

13

[7] J. Ashburner, Computational neuroanatomy, Ph.D. thesis, University CollegeLondon (2000).

[8] J. Ashburner, K. Friston, Unified segmentation, NeuroImage 26 (2005) 839–851.

[9] Y. Zhang, M. Brady, S. Smith, Segmentation of brain MR images througha hidden markov random field model and the expectation maximizationalgorithm, IEEE Transactions on Medical Imaging 20 (1) (2001) 45–57.

[10] B. Fischl, david Salat, E. Busa, M. Albert, M. Dietrich, C. Haselgrove,A. van der Kouwe, R. Kiliany, D. Kenedy, S. Klaveness, A. Mondtilo,N. Makris, B. Rosen, A. Dale, Whole brain segmentation: Automated labeling ofneuroanatomical structures in the human brain, Neurotechnique 2002 (January31) (1998) 341–355.

[11] J. Ren, B. Sneller, D. Rueckert, J. Hajnal, D. Hill, A comparison of tissue typelabelings of direct classification and segmentation propagation techniques forMR brain images, in: Proc. of Medical Image Understanding and Analysis ’05,2005, pp. 55–58.

[12] K. Pohl, W. Wells, A. Guimond, K. Kasai, M. Shenton, R. Kikinis, W. Grimson,S. Warfield, Incorporating non-rigid registration into expectation maximizationalgorithm to segment MR images, in: Proc. of the 5th Int. Conf. on MedicalImage Computing and Computer-Assisted Intervention, Vol. 2488 of LectureNotes in Computer Science, 2002, pp. 564–572.

[13] E. D’Agostino, F. Maes, D. Vandermeulen, P. Suetens, Non-rigid atlas-to-image registration by minimization of class-conditional image entropy, in: Proc.of the 7th Int. Conf. on Medical Image Computing and Computer-AssistedIntervention, Part I, 2004, pp. 745–753.

[14] K. Pohl, Prior information for brain parcellation, Ph.D. thesis, MassachusettsInstitute of Technology (2005).

[15] S. K. Warfield, M. Kaus, F. A. Jolesz, R. Kikinis, Adaptive template moderatedspatially varying statistical classification, in: Proc. of the 1st Int. Conf. onMedical Image Computing and Computer-Assisted Intervention, 1998, pp. 431–438.

[16] A. P. Z. Chris A. Cocosco, A. C. Evans, A fully automatic and robust brainMRI tissue classification method, in: Medical Image Analysis, Vol. 7, 2003, pp.513–527.

[17] M. Prastawa, J. H. Gilmore, W. Lin, G. Gerig, Automatic Segmentationof Neonatal Brain MRI., in: Proc. of the 7th Int. Conf. on Medical ImageComputing and Computer-Assisted Intervention, Part I, 2004, pp. 10–17.

[18] J. Hajnal, D. Hill, D. Hawkes, Medical Image Registration, CRC Press, 2001.

[19] A. P. Dempster, N. M. Laird, D. B. Rubin, Maximum likelihood from incompletedata via the EM-algorithm, Journal of the Royal Statistical Society 39 (1) (1977)1–38.

14

[20] A. E. A. Sled, J.G.; Zijdenbos, A nonparametric method for automaticcorrection of intensity nonuniformity in MRI data, IEEE Transactions onMedical Imaging 17 (1) (1998) 87–97.

[21] S. Smith, Fast robust automated brain extraction, Human Brain Mapping 17 (3)(2002) 143–155.

[22] Statistical Parametric Mapping, www.fil.ion.ucl.ac.uk/spm.

[23] Expectation Maximization Segmentation,www.medicalimagecomputing.com/EMS.

15