www.elsevier.com/locate/ynimg NeuroImage 34 (2007) 1612 – 1618

Optimum template selection for atlas-based segmentation Minjie Wu, a Caterina Rosano, b Pilar Lopez-Garcia, c Cameron S. Carter, c and Howard J. Aizenstein d,⁎ a

Department of Electrical and Computer Engineering, University of Pittsburgh, Pittsburgh, PA 15213, USA Department of Epidemiology, University of Pittsburgh, Pittsburgh, PA 15213, USA c Department of Psychiatry, University of California at Davis, Sacramento, CA 95817, USA d Department of Psychiatry, Western Psychiatric Institute and Clinic, University of Pittsburgh, 3811 O’Hara Street, Pittsburgh, PA 15213, USA b

Received 21 March 2006; revised 21 June 2006; accepted 31 July 2006 Available online 26 December 2006 Atlas-based segmentation of MR brain images typically uses a single atlas (e.g., MNI Colin27) for region identification. Normal individual variations in human brain structures present a significant challenge for atlas selection. Previous researches mainly focused on how to create a specific template for different requirements (e.g., for a certain population). We address atlas selection with a different approach: instead of choosing a fixed brain atlas, we use a family of brain templates for atlas-based segmentation. For each subject and each region, the template selection method automatically chooses the ‘best’ template with the highest local registration accuracy, based on normalized mutual information. The region classification performances of the template selection method and the single template method were quantified by the overlap ratios (ORs) and intraclass correlation coefficients (ICCs) between the manual tracings and the respective automated labeled results. Two groups of brain images and multiple regions of interest (ROIs), including the right anterior cingulate cortex (ACC) and several subcortical structures, were tested for both methods. We found that the template selection method produced significantly higher ORs than did the single template method across all of the 13 analyzed ROIs (two-tailed paired t-test, right ACC at t(8) = 4.353, p = 0.0024; right amygdala, matched paired t test t(8) > 3.175, p < 0.013; for the remaining ROIs, t(8) = 4.36, p < 0.002). The template selection method also provided more reliable volume estimates than the single template method with increased ICCs. Moreover, the improved accuracy of atlas-based segmentation using optimum templates approaches the accuracy of manual tracing, and thus is valid for automated brain imaging analyses. © 2006 Elsevier Inc. All rights reserved. Keywords: Atlas-based segmentation; Template selection

⁎ Corresponding author. Fax: +1 412 246 5880. E-mail address: [email protected] (H.J. Aizenstein). Available online on ScienceDirect (www.sciencedirect.com). 1053-8119/$ - see front matter © 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.neuroimage.2006.07.050

Introduction Atlas-based segmentation has become a standard method for automatically labeling regions of interest (ROIs) on MR brain images (Collins et al., 1995; Toga, 1999). In atlas-based segmentation, the standard template (atlas) is registered to the individual brain image by finding the optimal spatial transformation, and then mapping the anatomical information in the atlas onto the individual brain image. Normal individual variations in human brain structures present a fundamental and significant challenge for the atlas selection (Thompson et al., 2000). A common approach to address this issue is to construct a specialized atlas to meet specific requirements. For instance, the first widely used brain atlas, the Talairach template, was based on the brain of a 60-year-old female subject (Talairach and Tournoux, 1988). However, since this atlas is one particular brain, it does not fully reflect the variety of anatomical structures present in a population of normal brains. One approach is to generate probabilistic maps that retain information on the population variability (Mazziotta et al., 1995, 2001; Rademacher et al., 2001). An alternative approach is to construct a population-based atlas by averaging multiple co-registered brain images with a linear alignment to stereotaxic space (individual space → stereotaxic space), which represents group-specific features, but is blurred and lacks anatomical details (e.g., the atlas MNI305, Collins, 1994). To preserve the structural detail in the group image, a companion template (Colin27) was created by registering 27 high-resolution scans of a single individual to the blurred group MNI305 atlas (Holmes et al., 1998). An anatomical parcellation on Colin27 was also performed, which created the Automated Anatomical Labeling (AAL) atlas (Tzourio-Mazoyer et al., 2002). In a recent study we used an atlas-based segmentation method with the Colin27 atlas as the reference image and found lower anatomical labelling accuracy for the anterior cingulate cortex (ACC) than for the hippocampus (Wu et al., 2006). We suspected that the lower accuracy for the ACC was due to the normal inter-subject variability in the gyral folding pattern in

M. Wu et al. / NeuroImage 34 (2007) 1612–1618

this structure. It has been estimated that approximately 30–60% of the population have a paracingulate sulcus (PCS, Paus et al., 1996), a normal variant of the ACC, in which there is an additional gyral fold. Wide variability can also be observed for other cortical structures (Ono et al., 1990), which present a fundamental challenge in selecting the best template for automated anatomical labeling. A single brain template is unable to represent those regions with multiple normal anatomical variations (e.g., ACC) and thus the performance of atlas-based segmentation suffers. Most previous research involving atlas-based segmentation has used a single fixed template strategy (Dawant et al., 1999; Vemuri et al., 2003; Carmichael et al., 2005). Several other studies have demonstrated that specialized atlases are more appropriate for particular populations, such as children or the elderly (Thompson et al., 2001; Prastawa et al., 2005). In this study we take a different approach. Instead of choosing a fixed atlas such as Colin27 or a population-based atlas such as MNI305, we use a family of brain templates and for each subject we choose the ‘best’ template for the automated anatomical labeling. The intuition is that the variations in normal brain anatomy can be better represented as a small number of prototype atlases (e.g., presence or absence of paracingulate) rather than as a single average brain. For each subject, the template, which gives the optimum localized registration for a specific ROI (using a fully automated algorithm), is chosen as the optimum template for the automated anatomical labeling. This approach has previously been shown effective in atlas-based segmentation of bee brain images (Rohlfing et al., 2004). In the current study, this atlas selection technique was tested on two different human brain image data sets based on the automated anatomical labeling of multiple ROIs including right ACC, left and right amygdala, caudate, hippocampus, pallidum, putamen, and thalamus proper. For both data sets, the ROIs segmented using the optimum template selection method and the standard single template method were compared to the manual anatomical tracings respectively.

1613

passing through the AC. The cingulate and callosol sulci constituted the outer and inner boundary respectively. When a sulcus running parallel and superior to the cingulate sulcus was present, the paracingulate gyrus was included in the tracing. Inter-rater reliability for the right ACC manual tracings of the two raters was calculated using the intraclass correlation coefficient (ICC). The ICC for the right ACC was 0.97. To obtain intra-rater reliability, a subset of 5 MR images was retraced by the same rater after 3–4 weeks (mean 22.2 ± 3.4 days). The ICC for intra-rater reliability for the right ACC was 0.93. Data Set 2: Thirteen T1-weighted MR brain images from Internet Brain Segmentation Repository (IBSR) and their manual tracings were provided by the Center for Morphometric Analysis at Massachusetts General Hospital, available at http://www.cma.mgh. harvard.edu/ibsr/. Brain images were acquired from healthy control subjects. The 13 subjects from IBSR were at coronal orientation, with a matrix of 256 × 256 × 128, and had voxel sizes of 0.84 × 0.84 × 1.5 mm3 or 0.94 × 0.94 × 1.5 mm3 or 1 × 1 × 1.5 mm3. The T1-weighted images were ‘positionally normalized’ into the Talairach orientation (rotation only) and then processed with biasfield correction routines. The voxel size on the processed images was unchanged by the reorientation and biasfield correction. The manual tracings were all done by experts at the Center for Morphometric Analysis of Massachusetts General Hospital, Harvard Medical School. Iso-intensity contours were used in the manual tracing to define the ‘primary’ borders of anatomical structures (Filipek et al., 1994). The right ACC was manually traced on Data Set 1, and multiple ROIs including the left and right amygdala, caudate, hippocampus, pallidum, putamen, and thalamus proper were manually traced on Data Set 2. The manual tracings of both data sets served as gold standard tracings for evaluating the automatically segmented results. The manual anatomical tracings of the subject also served as the atlas when the subject brain image was used as the template. Registration method

Subjects and methods Subjects Two sets of data were used to evaluate the template selection approach. Both data sets have been previously described in greater detail: Data Set 1 (Wu et al., 2006) and Data Set 2 (http://www.cma.mgh.harvard.edu/ibsr/). Brief descriptions follow. Data Set 1: Nine subjects (6 male/3 female; mean age 24.3, range 20–32 years old; right-handed) participated. Scanning was done on a 1.5T GE CVi scanner with 3D SPGR (TR/TE = 5/25 ms, flip angle = 40°, FOV = 24 × 18 cm, slice thickness = 1.5 mm, matrix size = 256 × 192). The data were originally acquired with a voxel size of 0.94 × 0.94 × 1.5 mm3, and were resampled to a voxel size of 1 × 1 × 1 mm3. Two raters independently and manually classified the right ACC on each subject, which were used as the gold standard (i.e., best estimated) region mask. Right ACC tracings were made on serial coronal slices. The sagittal and axial views were used as a reference to outline the ACC. The posterior limit of the ACC was defined by a vertical line perpendicular to the anterior commissure–posterior commissure (AC–PC) plane and

Atlas-based segmentation labels the anatomical regions on individual images by registering the template image of the brain to the individual brain image. We refer to the registration method we use as the Automated Labeling Procedure (ALP). This is derived from the methods used by Chen (1999), which consists of preprocessing steps (including skull stripping and cropping) and a series of registration techniques including hierarchical registration (Unser et al., 1993) and demons based registration (Thirion, 1998). The ALP starts with 12 parameter affine registration to correct global differences such as orientation and brain size, and goes on with a grid-based piecewise linear registration for coarse alignment, then finally uses demons registration algorithm as a fine-tuning for a voxel-level spatial deformation. We have implemented this method using the registration library from the Insight Segmentation and Registration Toolkit (ITK, Yoo, 2004). The performance of this method was quantitatively compared to popular registration packages including Automated Image Registration (AIR, Woods et al., 1998) and Statistical Parametric Mapping (SPM, Ashburner and Friston, 1999), and was found to perform significantly better. Detailed descriptions and evaluations of the ALP are presented in Wu et al. (2006).

1614

M. Wu et al. / NeuroImage 34 (2007) 1612–1618

the maximum local NMI, is chosen as the locally optimized template for each subject’s ROI.

Template selection method The flow chart of the template selection algorithm is shown in Fig. 1. To select the optimum template for the subject and ROI, each subject is first registered to each template, which creates a warped image (template → subject); then, for each ROI, the registration accuracy of each template is evaluated between the warped image and subject image at a standard local region area O. This common area is formed as the disjunction image of the segmented ROIs Oi from all the templates. The local area O ¼ O1 [ O2 : : : [ On , with Oi as the segmented ROI using ith template for the subject. Normalized Mutual Information (NMI) is then used as the metric to measure the local registration performance: NM Iðx; yÞ ¼ ðHðxÞ þ HðyÞÞ=Hðx; yÞ

ð1Þ

where HðxÞ ¼ 

k X

Pðix Þlog2 ðPðix ÞÞ

ð2Þ

Pðiy Þlog2 ðPðiy ÞÞ

ð3Þ

i¼1

HðyÞ ¼ 

k X i¼1

Hðx; yÞ ¼ 

ky kx X X Pðix ; iy Þlog2 ðPðix ; iy ÞÞ

ð4Þ

ix ¼1 ix ¼1

where x is the cropped local ROI image of each warped image (each template → subject) and y is the cropped ROI image of the subject at the same local area; H(x), H(y) are the entropies and H(x,y) is the joint entropy of x and y. NMI describes the similarity of the warped image and the target image at a local ROI area, which also evaluates local registration accuracy and the performance of the template in the classification of the ROI. The template, which gives

Template evaluation methods Overlap ratio (OR) is used to quantify the region classification quality of each registration. OR is defined as the ratio of overlapping voxels to total voxels, as given below: overlap ratio ¼

ˆÞ volðB∩B ˆÞ volðB [ B

ˆ is the automatically where B is the manually traced ROI, and B segmented ROI on the subject. A perfect overlapping of the ˆ will manually traced ROI B and the automated labeled ROI B lead to an OR = 1, while less overlap will result in a smaller value (0 ≤ OR ≤ 1). For each data set, the segmentation performance of the individual atlas was evaluated based on a leave-one-out approach. Each of the 9 subjects from Data Set 1 was chosen to serve as an individual template and was registered to the remaining 8 subjects. Nine subjects out of the 13 subjects for Data Set 2 were randomly chosen as individual template and registered to the remaining 12 subjects. For each ROI, the performance of a single atlas was evaluated as the mean OR of the automatically labeled ROI against the manual traced ROI across multiple subjects using the same atlas. The average performance of the single template strategy was measured as the average OR across multiple templates. Additionally, the standard MNI (Montreal Neurological Institute) brain Colin27 (Holmes et al., 1998), which carries high anatomical details and has a high spatial resolution (1 × 1 × 1 mm3 voxel size), was also used as the template to segment the right ACC on Data Set 1 for comparison. In the template selection algorithm, a subject in data set 1 was warped to the 8 remaining atlases and the atlas with the best local registration accuracy was chosen; for consistency, the atlas was selected from a randomly chosen 9-subject subset of data set 2. The performance of the optimum template was measured as the mean OR across all the combinations of available templates for the ROI classification. The performances of the optimum template selection method and the single template strategy were also evaluated respectively by the absolute volume agreements between automatically segmented ROI and hand drawn ROI. Volume agreement between the automatically segmented ROI and the hand drawn ROI was measured in SPSS (Nie et al., 1970) using single measure ICC, with two-way mixed model and measures of absolute agreement. For the individual template case, the automatically labeled ROIs from each template were compared against the hand drawn ROIs across all subjects using ICC and the average ICC across all the template for segmenting the same ROI was used as the average performance of single template method. Results and discussion

Fig. 1. Template selection flowchart. The processing steps that constitute the template selection model, which is used to choose the optimum template from a family of templates (Tk) for the segmentation of ROI (R) on a subject (S).

The evaluations on the automated anatomical labeling results from both methods (single template and optimum template) were compared for each data set. Both methods used exactly the same pathway for the registration, as well as the same thresholds (to

M. Wu et al. / NeuroImage 34 (2007) 1612–1618

remove the edges of the automated anatomical segmentations). The only difference is that they used different atlas selection strategies. We found that for most of the ROIs the optimum template produced significantly and consistently better classification results compared to the single template method. Data set 1 As predicted, atlas-based segmentation with the template selection method produced significantly better mean ORs for right ACC classification across the 9 subjects than with any single template. The mean ACC OR when registering with a single template ranged from 42.7% to 52.7% (mean OR 49.5%), and the mean OR using the standard template MNI Colin27 was 47.3%; while the mean OR with the optimum template selection method was 54.7% (8 templates); this method reached an OR of 57.7% when an optimized template subset was used (3 templates, Fig. 2). As seen in Fig. 2, the performance of the optimum template selection method was better than the performance of any of the template candidates or the standard MNI template Colin27. A twotailed paired t test showed that the registration result from optimum template (8-template case) based method was significantly better than the results from single templates at t(8) = 4.353, p = 0.0024. One subject image, 3 template images and the corresponding NMI results from the template selection model are shown in Fig. 3. It can be noted from the figure that the target subject has a paracingulate sulcus (in red), and that template 2, which has a similar paracingulate sulcus (in red) was automatically selected as the optimum template to segment the right ACC on the subject using the maximum NMI between the warped image and the target subject image at local ACC areas. The OR of the automatically labeled right ACC from the 3 templates against the manual tracing on the target image was also calculated and compared to validate the performance of the template selection model. It is important to determine how many templates are needed to achieve robust automated anatomical labeling of certain ROIs with the template selection method. In the previous test, for each subject the remaining 8 brain images were used as template candidates for the classification of right ACC. To test how the performance of the template selection method changes with the number of templates, and to decide how many templates are sufficient for the

Fig. 2. In the anatomical classification of right ACC, mean overlap ratios (OR) are compared between individual template strategy (average of the mean OR over the 9 templates) and the optimal template method with 8 templates or with the optimized 3 templates. Error bars were calculated as standard error of the difference between the mean OR of the optimal template and the individual templates. The OR for the Colin template is shown for comparison.

1615

classification of specific ROI, we tried the template selection method on this dataset using all subsets of the 8 different templates (from 1 to 8 templates in each subset). The performance of the template selection method with a particular number of template candidates was estimated as the average OR across all subjects for all subsets of templates of that cardinality. For example, the performance of the template selection method with 2 templates was estimated as the mean OR across the 7 subjects (excluding the 2 templates) and across all possible template combinations (C29 = 36). The mean ORs are plotted in Fig. 4 against the number of templates used. As expected, as the number of templates increased, the average performance of the template selection method improved. For each subset of templates (N templates), the performance was estimated as the average OR across subjects (9-N). We observed that among the template subsets of the same cardinality, the template set with the widest anatomical variations performed better than other combinations of templates. For example, in the automated classification of right ACC, we discovered there was a sufficient set of three templates, which included three prototypes: one template with a paracingulate sulcus, one with a thick anterior cingulate cortex, and the third with a thin anterior cingulate cortex, which had the best performance and achieved an average OR of 0.577 (similar to that of the 8 template case), as shown in Fig. 3. This suggests that in addition to template number, template variability is also important, such that with an appropriately variable set of templates, a fewer number of templates are sufficient for high accuracy. Data set 2 The automated anatomical labeling results for multiple regions (left and right amygdala, caudate, hippocampus, pallidum, putamen and thalamus proper) on the IBSR dataset using the template selection method and the single template method were compared to manual tracings respectively using OR and volume agreement. For most of the ROIs, the template selection method consistently provided more reliable region classification than the single template method. The mean percent ORs and ICCs between the estimated volumes and manual traced volumes for both methods are shown in Fig. 5. For data set 2, the template selection method gave a higher mean OR than the single template method for all ROIs; the OR percent increase ranged from 4.4% to 13.1% (mean increase 8%). The differences in ORs were highly significant for all ROIs (for right amygdala, matched pair t(8) > 3.175, p < 0.013; for the remaining ROIs, t(8) > 4.36, p < 0.002). Also the template selection method provided a more reliable volume estimate. The ICCs between the estimated volumes and the manual tracings from the template selection method were higher than the results from single template for 11 of 12 ROIs. The changes in the ORs were consistent with the increased ICCs from the volume agreement. For example, the large increase of ORs in regions like left caudate (11.1% increase), right caudate (12.0%), and right pallidum (13.1%) corresponded to more improved ICCs in left caudate (ICC 0.95), right caudate (ICC 0.93) and right pallidum (ICC 0.91). For all of the ROIs except for left and right amygdala and right putamen, the template selection method produced very reliable ICCs of volume estimates, comparable with ICCs of interrater manual tracings, which is essential to MR volumetric studies on such ROIs. For instance, the template selection method produced ICCs of 0.95 for the left caudate, and 0.93 for the right

1616

M. Wu et al. / NeuroImage 34 (2007) 1612–1618

Fig. 3. For data set 1 the optimum template selection model selects the best template from a family of 8 templates to segment the right ACC on the target image. Three of the templates and the target subject images are shown here; the hand-drawn ACC on the subject and templates are also displayed in color (cingulate in blue and paracingulate in red). Also shown are the normalized mutual information (NMI) calculated by comparing the warped template with the target image, and the overlap ratios (ORs), calculated by comparing the automated labeled result with the manual tracing.

caudate, while previous research reports inter-rater reliability of 0.94 for the left caudate and 0.95 for the right caudate (Venkatasubramanian et al., 2003). To illustrate the absolute agreement between the estimated volumes and the manual volumes, the estimated volumes of four ROIs on data set 2 from both methods were plotted against the manual volumes in Fig. 6. As Fig. 6 illustrates, the accuracy of volume estimates improved considerably by using the optimum template method. Overall, the optimum template method provided better volume estimates than the single template method. However, for the left and right amygdala, neither method gave a good estimate. This is because the registration failed to provide an accurate region classification with any of the template candidates.

We found that the template selection model produced significantly better ORs and more reliable volume estimates in the analyzed multiple ROIs than the single template strategy. The segmentation

Conclusion In this study, multiple prototype atlases were used to address the normal brain anatomy variations in the atlas-based segmentation of MR brain images. The template selection algorithm uses normalized mutual information to choose the template (from a family of templates) that gives the best local registration accuracy. This template selection model is of special use to those regions with high variability across subjects such as cortical structures (Ono et al., 1990), where a single template can not readily capture the variability.

Fig. 4. The performance of the multiple template method with different number of templates. The mean ORs across all the combinations of samenumber templates were plotted against the number of templates used.

Fig. 5. Comparison of the reliability of the automated ROIs using the template selection method and using a single template. Key: LA—left amygdala, RA—right amygdala, LC—left caudate, RC—right caudate, LH—left hippocampus, RH—right hippocampus, LPa—left pallidum, RPa—right pallidum, LPu—left putamen, RPu—right putamen, LT—left thalamus proper, RT—right thalamus proper. Top: mean percent OR comparion. Bottom: the intraclass correlation coefficients (ICCs) of volume agreement for both methods.

M. Wu et al. / NeuroImage 34 (2007) 1612–1618

1617

Fig. 6. Region classification performance of the optimum template method and the single template method. The absolute voxel numbers of segmented ROIs from both methods are compared to manual segmentation respectively using the linear regression model. Four ROIs (left and right caudate, and left and right thalamus proper) were analyzed for 13 subjects. Subject 6 was used as the atlas in the single template method; the optimum template was chosen from a 9-subject subset.

accuracy improved with the increasing number of the templates used in the template selection method. In addition to the template number, the anatomical variability within the used templates is also important, such that a fewer number of templates with appropriate anatomical variations are sufficient to achieve high accuracy in the atlas-based segmentation; in the case of right ACC, we found a set of 3 prototype templates with wide anatomical variations, which performed better than other template combinations. This higher registration accuracy with the template selection model is achieved at the cost of higher computation load. Multiple non-rigid registrations are required in order to evaluate the performance of multiple templates. Also although this method produced improved anatomical classification accuracy for all the analyzed ROIs, it did not give satisfactory region estimates for the left and right amygdala. Alternative classification methods (perhaps using feature-based registration) should be used to improve the automatical labeling of the small and difficult to segment regions such as the left and right amygdala. Also more advanced method may be needed to evaluate the performance of the templates at such regions. The number of available atlases limits the registration accuracy. More atlas candidates may bring higher registration accuracy, but with extra computation load, since in order to evaluate the templates we need to register each template to the target subject. Also the number of atlases needed is related to the normal anatomic variations of the region to be segmented. Different ROIs may require different

numbers of atlas prototypes for the automated anatomical classification. In this study, we discussed the possible atlas prototypes for the anterior cingulate cortex (thin ACC, thick ACC, and with a paracingulate sulcus). Further research is needed to explore the normal variations of the brain anatomy. Multiple templates are needed in the template selection method. The templates can either be manually traced locally by experts, or downloaded from a public database. There are many manually labeled atlases available online, such as the IBSR dataset used in this paper (http://www.cma.mgh.harvard.edu/ibsr/), which has 18 high-resolution T1-weighted MR image data with expert segmentations of 43 individual structures. This method chooses the best atlas from a family of atlases for each subject and ROI, and is independent of the registration techniques. In our study, we used the deformable automated labeling pathway for the inter-subject registration, and it can be easily accommodated into alternate available pathways such as Automated Image Registration (AIR) or Statistical Parametric Mapping (SPM). Acknowledgments This research was supported by NARSAD, The Pittsburgh Foundation, Burroughs Wellcome Translational Scientist Award (CSC), and NIMH grants K02-MH064190, K23 MH064678 P30 MH052247 and NIA grant P30 AG024827, Pittsburgh Claude D.

1618

M. Wu et al. / NeuroImage 34 (2007) 1612–1618

Pepper Older Americans Independence Center. We thank the IMAGe (Imaging Methods and Analysis in Geriatrics) group and the Clinical Cognitive Neuroscience laboratory at the University of Pittsburgh for their assistance. References Ashburner, J., Friston, K.J., 1999. Nonlinear spatial normalization using basis functions. Hum. Brain Mapp. 7, 254–266. Carmichael, O.T., Aizenstein, H.J., Davis, S.W., Becker, J.T., Thompson, P.M., Meltzer, C.C., Liu, Y., 2005. Atlas-based hippocampus segmentation in Alzheimer's disease and mild cognitive impairment. NeuroImage 27 (4), 979–990. Chen, M., 1999. 3-D Deformable Registration Using A Statistical Atlas with Applications in Medicine. Carnegie Mellon University. Collins, D.L., 1994. 3D Model-Based Segmentation of Individual Brain Structures from Magnetic Resonance Imaging Data. McGill University, Montreal. Collins, D.L., Holmes, C.J., Peters, T.M., Evans, A.C., 1995. Automatic 3-D model-based neuroanatomical segmentation. Hum. Brain Mapp. 3, 190–208. Dawant, B.M., Hartmann, S.L., Thirion, J.-P., Maes, F., Vandermeulen, D., Demaerel, P., 1999. Automatic 3-D segmentation of internal structures of the head in MR Images using a combination of similarity and freeform transformations: Part I, methodology and validation on normal subjects. IEEE Trans. Med. Imag. 18 (10), 909–916. Filipek, P.A., Richelme, C., Kennedy, D.N., Caviness Jr., V.S., 1994. The young adult human brain: an MRI-based morphometric analysis. Cereb. Cortex 4 (4), 344–360. Holmes, C.J., Hoge, R., Collins, L., Woods, R., Toga, A.W., Evans, A.C., 1998. Enhancement of MR images using registration for signal averaging. J. Comput. Assist. Tomogr. 22 (2), 324–333. Mazziotta, J.C., Toga, A.W., Evans, A., Fox, P., Lancaster, J., 1995. A probabilistic atlas of the human brain: theory and rationale for its development. The International Consortium for Brain Mapping (ICBM). NeuroImage 2 (2), 89–101. Mazziotta, J., Toga, A., Evans, A., Fox, P., Lancaster, J., Zilles, K., Woods, R., Paus, T., Simpson, G., Pike, B., et al., 2001. A probabilistic atlas and reference system for the human brain: International Consortium for Brain Mapping (ICBM). Philos. Trans. R. Soc. Lond., B Biol. Sci. 356 (1412), 1293–1322. Nie, N.H., Bent, D.H., Hull, C.H., 1970. SPSS: Statistical Package for the Social Sciences. McGraw-Hill, New York. xx, 343 pp. Ono, M., Kubik, S., Abernathey, C.D., 1990. Atlas of the Cerebral Sulci. Stuttgart, New York: G. Thieme Verlag; Thieme Medical Publishers. xiv, 218 pp. Paus, T., Tomaiuolo, F., Otkay, N., MacDonald, D., Petrides, M., Atlas, J., Morris, R., Evans, A.C., 1996. Human cingulate and paracingulate sulci: pattern, variability, asymmetry, and probabilistic map. Cereb. Cortex 6, 207–214.

Prastawa, M., Gilmore, J.H., Lin, W., Gerig, G., 2005. Automatic segmentation of MR images of the developing newborn brain. Med. Image Anal. 9 (5), 457–466. Rademacher, J., Morosan, P., Schormann, T., Schleicher, A., Werner, C., Freund, H.-J., Zilles, K., 2001. Probabilistic mapping and volume measurement of human primary auditory cortex. NeuroImage 13 (4), 669–683. Rohlfing, T., Brandt, R., Menzel, R., Maurer, C.R.J., 2004. Evaluation of atlas selection strategies for atlas-based image segmentation with application to confocal microscopy images of bee brains. NeuroImage 21 (4), 1428–1442. Talairach, J., Tournoux, P., 1988. Co-planar Stereotaxic Atlas of the Human Brain: 3-Dimensional Proportional System: An Approach to Medical Cerebral Imaging. Thieme Medical Publishers, Stuttgart. 122 pp. Thirion, J.P., 1998. Image matching as a diffusion process: an analogy with Maxwell's demons. Med. Image Anal. 2, 243–260. Thompson, P.M., Woods, R.P., Mega, M.S., Toga, A.W., 2000. Mathematical/computational challenges in creating deformable and probabilistic atlases of the human brain. Hum. Brain Mapp. 9 (2), 81–92. Thompson, P.M., Mega, M.S., Woods, R.P., Zoumalan, C.I., Lindshield, C.J., Blanton, R.E., Moussai, J., Holmes, C.J., Cummings, J.L., Toga, A.W., 2001. Cortical change in Alzheimer's disease detected with a diseasespecific population-based brain atlas. Cereb. Cortex 11 (1), 1–16. Toga, A.W., 1999. Brain Warping. Academic Press, San Diego. Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., Mazoyer, B., Joliot, M., 2002. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. NeuroImage 15, 273–289. Unser, M., Aldroubi, A., Gerfen, C.R., 1993. A Multiresolution image registration procedure using spline pyramids. Proceedings of the SPIE Conference on Mathematical Imaging: Wavelet Applications in Signal and Image Processing; 1993 July 15–16; San Diego CA, USA, pp. 160–170. Vemuri, B.C., Ye, J., Chen, Y., Leonard, C.M., 2003. Image registration via level-set motion: applications to atlas-based segmentation. Med. Image Anal. 7 (1), 1–20. Venkatasubramanian, G., Gangadhar, B.N., Jayakumar, P.N., Janakiramaiah, N., Keshavan, M.S., 2003. Striato-cerebellar abnormalities in nevertreated schizophrenia: evidence for neurodevelopmental etiopathogenesis. Ger. J. Psychiatry 6 (1), 1–7. Woods, R.P., Grafton, S.T., Watson, J.D., Sicotte, N.L., Mazziotta, J.C., 1998. Automated image registration: II. Intersubject validation of linear and nonlinear models. J. Comput. Assist. Tomogr. 22 (1), 153–165. Wu, M., Carmichael, O., Lopez-Garcia, P., Carter, C.S., Aizenstein, H.J., 2006. Quantitative comparison of the AIR, SPM, and a fully deformable model for atlas-based segmentation of functional and structural MR images. Hum. Brain Mapp. 27 (9), 747–754. Yoo, T., 2004. Insight into Images: Principles and Practice for Segmentation, Registration, and Image Analysis. AK Peters Ltd, Wellesey, MA.