- Email: [email protected]
The Talairach coordinate of a point in the MNI space: how to interpret it
www.elsevier.com/locate/ynimg NeuroImage 25 (2005) 408 – 416
The Talairach coordinate of a point in the MNI space: how to interpret it Wilkin Chau* and Anthony R. McIntosh The Rotman Research Institute, Baycrest Centre for Geriatric Care, University of Toronto, 3560 Bathurst Street, Toronto, Ontario, Canada M6A 2E1 Received 18 March 2004; revised 2 December 2004; accepted 6 December 2004
To perform group studies using functional imaging data, the individual brain images are usually transformed into a common coordinate space. The two most widely used spaces in the neuroscience community are the Talairach space and the Montreal Neurological Institute (MNI) space. The Talairach coordinate system has become the standard reference for reporting the brain locations in scientific publication, even when the data have been spatially transformed into different brain templates (e.g., MNI space). When expressed in terms of individual subjects, the mapping of a coordinate in MNI space to the Talairach space generates distinct coordinates for different subjects. In this paper, we describe two approaches to derive the Talairach coordinates from the MNI space, which is based on the ICBM152 template from the International Consortium of Brain Mapping. One approach is the Talairach Method of Piecewise Linear Scaling (TMPLS) as implemented in the AFNI software package; and the other is a template-matching approach using the linear transformation in SPM99. The uncertainty measurements of the mapping results are presented. This may allow researchers to better interpret results reporting in the Talairach coordinates obtained from the MNI space. This study also examines the discrepancy between the derived Talairach coordinates and those obtained from the mni2tal script, a tool commonly used by the neuroimaging community. Large discrepancies are found in the inferior regions, superior frontal and occipital regions. D 2005 Elsevier Inc. All rights reserved. Keywords: Talairach space; MNI space; Spatial normalization; Brain mapping
Introduction Functional neuroimaging studies typically rely on group studies to identify the most consistent pattern of activation associated with a given cognitive operation. The neuroanatomical difference
* Corresponding author. Fax: +1 416 785 2862. E-mail address: [email protected] (W. Chau). Available online on ScienceDirect (www.sciencedirect.com). 1053-8119/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.neuroimage.2004.12.007
between individuals make it important to accurately co-register the brain images of multiple subjects. One approach is based on anatomical landmarks, which requires manually identification of the region of interest (ROI) in each individual brain. This is very costly in terms of time and sometimes ROIs may cross functional boundaries. Most researchers use an alternative approach that based on transformation to a particular brain atlas coordinate system. This approach usually involves an automated process to spatially transform the individual brain image into the coordinate space. However, it does not guarantee that an identical point in the space corresponds to the same anatomical feature for all subjects. The two most widely used spaces in the neuroscience community are the Talairach space (Talairach and Szikla, 1967; Talairach and Tournoux, 1988) and the Montreal Neurological Institute (MNI) spaces (Evans et al., 1993). The Talairach space is based on a stereotaxic atlas of the human brain published by Talairach and Tournoux (Talairach and Szikla, 1967; Talairach and Tournoux, 1988). They identified the anatomical features from the atlas and created a coordinate system related to anatomical landmarks. The Talairach coordinate space has its origin defined at the anterior commissure (AC), with x- and y-axes on a horizontal plan and z-axis on a vertical plane. The yaxis is defined by the line connecting the most superior of AC and the most inferior of the posterior commissure (PC); the x-axis is defined by the line that passes through the AC point and orthogonal to the AC-PC line; whereas the z-axis is the line that passes through the interhemispheric fissure and the AC point. Given a 3-D coordinate in the Talairach space, the anatomical labels can be obtained manually through inspection of the atlas. To accommodate the differences between individual brains, Talairach and Tournoux defined a proportional grid system to align an individual brain to the atlas. The procedure involves dividing the brain into 12 regions by using one horizontal plane, passing through the x- and y-axes, and three vertical planes. Two vertical planes are parallel to the x-axis, one passing through the z-axis and another through the PC point, and one is parallel to the y-axis, dividing the left and right hemispheres. Each region is then scaled, separately for each direction, to match the atlas. This piecewise linear scaling method provides a simple means of
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
converting an individual brain to the Talairach space; however, there is no guarantee that the transformed brain would completely match either the shape or the anatomical features of the Talairach atlas. This linear scaling method performs poorly in terms of matching the anatomical landmark locations to the atlas, especially cortical regions, compared to other methods that use nonlinear transformation. The major criticism about the Talairach atlas is that it was created based on the postmortem brain of single subject, which is not a good representation of the neuroanatomy for the general population. To allow better representation of average neuroanatomy, the MNI created an average brain template based on the MRI scans from several hundred individuals. The first template, known as the MNI305, was created in two steps. The first step was to obtain the average brain of 241 brains; each of them had been reoriented and scaled to match a set of manually selected anatomical landmarks to those of Talairach atlas. The second step was to create the MNI305 template by averaging the 305 normal MRI scans, which had been normalized with a linear transformation matrix to match the average brain created in the first step. Later, a template using the 152 brains normalized to MNI305 template was created. The International Consortium of Brain
409
Mapping (ICBM) has adopted the template as an international standard, known as the ICBM152. This template is used by several functional imaging analysis packages, such as SPM99 (Wellcome Department of Cognitive Neurology, London, UK; http://www.fil. ion.ucl.ac.uk/spm) and FSL (Image Analysis Group, FMRIB, Oxford, UK; http://www.fmrib.ox.ac.uk/fsl). The MNI templates are not perfectly matched with the Talairach standard brain due to large differences in brain shape and size between the two templates. Moreover, individual differences also prompted the ICBM to develop a probabilistic atlas of the human brain (Mazziotta et al., 1995). Since the Talairach coordinate system has become the standard reference for reporting the brain locations in scientific publication, researchers often need to report their findings using this system, even though their data were analyzed in different coordinate space, such as the ICBM152. There is no simple way to transform multiple subject data from the MNI space to the Talairach space. It is quite possible that the coordinate location in MNI space of two subjects would map to different points of Talairach space (Fig. 1a). The discrepancy becomes an issue when the data are analyzed in the MNI space but the results are reported using the Talairach space (Brett et al., 2002).
Fig. 1. (a) Mapping from the SPM99 MNI space to the Talairach space. A point in the MNI space (left) is associated with two different locations of two of the subjects (middle). The locations in the brains correspond to two distinct Talairach coordinates. (b) The schematic diagram of three different ways to obtain the Talairach coordinates of a point in the MNI space. (A) Converted from the MNI space using mni2tal script; (B) using TMPLS approach; (C) using the templatematching approach.
410
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
In this study, we investigate different approaches of deriving the Talairach coordinates from the MNI coordinates and report the uncertainty measurements for the approaches. This study focuses on the coordinate mapping approach, in which the transformation of each individual brain to the common space is independent to the underlying anatomy. More specifically, we studied the results of converting the MNI coordinates, obtained from transforming individual brain images to the ICBM152 template with the SPM99 software, to the Talairach coordinates using two different approaches: (1) the Talairach Method of Piecewise Linear Scaling (TMPLS) using the AFNI software (Cox, 1996; http://afni.nimh. nih.gov/afni); and (2) the linear transformation in SPM99 to match a template that matched the Talairach coordinate space. Finally, we compared the derived Talairach coordinates with those obtained from the mni2tal script (http://eeg.sourceforge.net/mridoc/mri_toolbox/mni2tal.html), a tool commonly used to map the MNI coordinates to the Talairach coordinates (Fig. 1b). This study demonstrates the uncertainty of the transformation from the MNI coordinates to the Talairach coordinates for multiple subject data, showing the spatial extent of the distribution of points in the Talairach space converted from a given MNI coordinates using the above two approaches. Our findings provide a guide for interpretation of the Talairach coordinates obtained from the MNI space. It is especially useful for researchers who conduct meta-analysis using data from both MNI space and Talairach space.
Materials and methods Subjects The MRI scans of 56 subjects (23 females) were obtained from several different fMRI experiments in our institute. Subjects were between the ages of 20 and 38 years (mean age 28 years). All subjects had no history of neurological disorders. Informed consent was obtained from each subject after the nature of each study had been explained. MRI data acquisition All MRI data were acquired with a 1.5-T MR scanner (Signa, General Electric Medical Systems, Waukesha, WI). A standard quadrature transmit/receive birdcage head coil was used. Head restraint was provided using a vacuum pillow (Vac Fix, Par Scientific, Inc.). Anatomical T1-weighted MRI was acquired using
three-dimensional fast spoiled gradient echo imaging technique [124 axial slices 1.4 mm thick, field-of-view (FOV) = 22 22 cm, flip angle/TE/TR = 358/6 ms/16 ms, 1 NEX, 256 256 acquisition matrix]. Data processing SPM99 was used to transform the anatomical data into the MNI space, using the T1 template of ICBM152 provided with the software. The software uses the minimization algorithms to determine the transformation parameters that allow the best match the voxel intensity levels of the template. Two types of transformation were considered: linear and nonlinear. The linear transformation has 12 parameters, which define the translation, rotation, scaling, and shearing operations. The nonlinear transformation is based on a set of cosine basis functions (Ashburner and Friston, 1999). The parameters were set to the default for SPM 99 as follows: the number of basis function was set to 7 8 7, the number of iterations was 13, and the medium regularization was used (i.e., with the value of 0.01). A set of grid points, separated by 4 mm in each direction, in the MNI space were defined as the sample of locations of interest to be compared. To derive the Talairach coordinates for each coordinate in the MNI space, the locations corresponding to the MNI coordinates were identified in each individual brain image. These locations were obtained by extracting the transformation matrices from the SPM spatial normalization operation. In SPM99, the spatial normalization process first determines the coefficients for the affine transformation matrix, and those for the basis functions when nonlinear transformation is performed. These matrices define that transformation from source space (the subject’s brain image in our case) for each sample point to the target space (the MNI space). The corresponding location of individual brain image to each MNI coordinate was obtained by extracting the information from the SPM script of bspm_write_sn.m,Q which contains the source location information for each point in the target space. Once the location in the subject brain was determined, transformations were performed to map the location into the Talairach space. The Talairach Method of Piecewise Linear Scaling (TMPLS) has been adapted in AFNI. Another approach is to map the individual brain to a template that conforms to Talairach coordinate space. Both approaches, the TMPLS and template normalization processes, were used in this study. For the TMPLS approach, the anatomical landmarks of the anterior commissure (AC) superior edge, AC posterior margin, posterior commissure (PC), and two mid-sagittal points, as well as the bounding box of
Table 1 Talairach coordinates of the eight sample locations in the SPM99 MNI space MNI coordinates
64, 16, 12 60, 24, 4 56, 60, 8 48, 72, 40 48, 12, 44 8, 56, 4 12, 24, 64 12, 92, 0
MNI2TAL coordinates
63, 16, 10 59, 23, 5 55, 57, 9 48, 72, 33 48, 10, 41 8, 54, 6 12, 26, 58 12, 89, 4
Centroid of the clusters using TMPLS
Centroid of the clusters using Talairach template
Linear
Linear
58, 18, 14 (3.29) 56, 19, 9 (3.73) 51, 58, 7 (3.84) 43, 72, 35 (4.94) 44, 5, 44 (4.57) 8, 52, 11 (4.03) 10, 27, 60 (4.79) 9, 87, 3 (4.90)
Nonlinear 60, 17, 13 (3.35) 58, 21, 7 (4.21) 53, 57, 9 (4.12) 43, 74, 35 (5.05) 44, 7, 44 (4.79) 9, 50, 11 (4.46) 11, 25, 59 (4.72) 10, 89, 6 (5.08)
60, 15, 17 (0.63) 56, 24, 11 (0.65) 52, 56, 2 (0.83) 44, 69, 41 (0.82) 45, 11, 46 (0.49) 6, 53, 11 (0.59) 10, 24, 63 (0.60) 10, 87, 4 (0.35)
The mean distance (in millimeter) between the cluster centroid and the points within the cluster is shown in bracket.
Nonlinear 60, 15, 18 (2.10) 56, 23, 11 (2.07) 53, 56, 3 (2.06) 44, 69, 41 (2.25) 44, 11, 46 (2.01) 7, 52, 11 (1.68) 10, 24, 63 (2.19) 9, 88, 5 (2.20)
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416 Fig. 2. Clusters in the Talairach space obtained by the TMPLS approach. The corresponding Talairach coordinates of 56 subjects for eight locations in the MNI space. (a) Cluster points from a point of MNI space obtained from a linear transformation. Left: The lateral and medial views of the cluster points. Right: The distribution of cluster points of the superior temporal location; red dot: the cluster centroid, green dot: the output of MNI2TAL script. (b) Same as panel a, but for the MNI data obtained from a nonlinear transformation.
411
412 W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
Fig. 3. Same as Fig. 2, but for the template-matching approach.
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
the brain (i.e., dorsal–ventral, anterior–posterior, and lateral extremes), were manually identified using the AFNI software. From these landmarks, the necessary rotation and linear scaling to transform the brain into the Talairach space is performed. For the template-matching approach, a Talairach template was created using the following procedure: first, each of the 56 brains was transformed into the stereotaxic atlas space using AFNI described above; second, a mean volume was created based on the transformed brains obtained in the step one; and third, the mean volume was smoothed with a Gaussian filter (FWHM = 5 mm) to create the Talairach template. Once the template was created, an affine transformation is performed to map each subject brain to the template using SPM99. Instead of deriving the Talairach coordinates through the subject space, a simple MATLAB script, called bmni2tal,Q written
413
by Matthew Brett (MRC Cognition and Brain Sciences Unit, Cambridge, England), can be used to determine the approximate Talairach coordinates for a point in the MNI space. The conversion is based on two linear transformation matrices, one for the regions above the AC point and the other for the regions below. Unlike the TMPLS and template-matching approaches, the bmni2talQ provides a single Talairach coordinate for each point in the MNI space, which is independent of the subject. Uncertainty measurements Due to difference between the subject brains, the conversion from the MNI space to the subject space, and then to the Talairach space is unique for each subject. A coordinate in the MNI space may be mapped to different Talairach coordinates for different
Fig. 4. The uncertainty measurement of the mapping from the MNI coordinates to the Talairach coordinates using the TMPLS approach. The brain slices are in the MNI space, with z coordinate shown below each slice. The color code represents the uncertainty measurement, which is related to the size of the cluster in the Talairach space corresponding the MNI coordinate. Similar pattern is found when using the template-matching approach.
414
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
subjects. With multiple subjects each point in the MNI space is associated with a cluster of points in the Talairach space. To determine the best Talairach coordinate to represent a given point in the MNI space, the centroid of a cluster points in Talairach coordinates associated with the MNI coordinate was calculated. The mean Euclidean distance between the centroid and the points within the cluster was used to quantify the spatial extent of the cluster. By examining the distribution of the mean distance in the brain, the uncertainty of the mapping from the MNI space to the Talairach space of any given region was determined.
Results Distribution of points in the Talairach space To demonstrate the distribution of the points in the Talairach space, eight points in the MNI space were selected to cover the medial and lateral regions of the brain (Table 1). The Talairach coordinates were obtained by the three approaches described above, namely, (1) the TMPLS approach; (2) affine transformation to match a Talairach template; and (3) the mni2tal scripts. The clusters associated with each MNI coordinate for the first two approaches are shown in Figs. 2 and 3. For the TMPLS approach, the spatial extents of these eight clusters are very similar. The range of the mean distance within each cluster is 3.29–4.94 mm (mean = 4.26 mm) when the linear transformation was used to convert the subject space to the MNI space, and 3.35–5.08 mm (mean = 4.48 mm) for the nonlinear transformation. For the template approach, the spatial extents of the clusters are consistently smaller than those by linear scaling. The range of the mean distance within each cluster is 0.35–0.83 mm (mean = 0.64 mm) for the linear conversion, and 1.68–2.77 mm (mean = 2.15 mm) for the nonlinear one. Uncertainty measurement of the MNI to Talairach space transformation The mean distance between the cluster centroid and each point within the cluster, which is the Talairach coordinates for each subject, is used to measure the uncertainty. For the TMPLS approach, there are two regions corresponding to the largest uncertainty, with values above 5 mm. One is located in the superior frontal region while the other is a band running from the superior occipital to the inferior frontal regions (Fig. 4). The distributions of the uncertainty patterns are very similar for both linear and nonlinear transformations from subject space to the MNI space; the difference of these two patterns is always less than 1 mm. The mean value of the uncertainty is 4.06 and 4.31, respectively. The differences of the Talairach coordinates found associated with the two transformations are always less than 3 mm. From the histograms of the uncertainty values (not shown), more points in the MNI space are found to have values smaller than 4 mm for the linear transformation compared to the nonlinear one. The number of points that have values larger than 4 mm is similar for both transformations. For the template-matching approach, no clear difference in pattern is found associated with the linear and nonlinear transformations. For the linear transformation to the MNI space, the uncertainty values are in the range of 0.3–1.16 with mean = 0.68. For the nonlinear case, the uncertainty values are in the range of
1.45–2.88 with mean = 2.10. Even with different uncertainty values, the differences of the Talairach coordinates found associated with these two transformations are always less than 3 mm, which is similar to those of the TMPLS approach. Performance of the MNI2TAL The performance of the mni2tal program was evaluated through computing the Euclidean distance between the Talairach centroids of the clusters obtained above and the mni2tal output. As an example, the centroid of the sample cluster in the superior temporal region and the corresponding Talairach coordinate found by mni2tal can be found in the scatter plots of Figs. 2 and 3, and the discrepancy values are shown in Table 2. The discrepancy between the centroids and the mni2tal output points is shown in the Fig. 5. The discrepancy between the TMPLS and the mni2tal output is in the range of 3.0–9.5 mm (mean = 6.27 mm) for the linear transformation, and 3.8–9.5 mm (mean = 6.54 mm) for the nonlinear transformation; the discrepancy between the templatematching and the mni2tal output is in the range of 2.7–10.6 mm (mean = 6.14 mm) and 3.6–10.8 mm (mean = 6.69 mm) for the linear and nonlinear transformation, respectively. In general the regions with large uncertainty values are most likely associated with large discrepancy values within each approach. Major discrepancy is observed from the superior frontal region and most of the points with z value b12 mm in the MNI space have discrepancy larger than 5 mm (Fig. 5). Similar pattern of discrepancy is found for both linear and nonlinear transformations from subject space to the MNI space. The discrepancy for the template matching is relatively larger compared to the TMPLS, especially for the inferior regions. For the TMPLS approach, the discrepancy difference between the linear and the nonlinear transformations is in the range of 1.3–1.56 mm (mean = 0.28 mm). For the template-matching approach, the discrepancy difference between the linear and the nonlinear transformations is in the range of 1.3–2.08 mm (mean = 0.54 mm).
Discussion The Talairach and Tournoux atlas has become the de facto standard for localization of in neuroimaging data sets, despite being based on a single postmortem brain. Better representation of the averaging neuroanatomy is achieved through image templates from averaging multiple individuals. The templates developed by
Table 2 Discrepancy between the MNI2Tal and the empirical-derived Talairach coordinates MNI coordinates
64, 16, 12 60, 24, 4 56, 60, 8 48, 72, 40 48, 12, 44 8, 56, 4 12, 24, 64 12, 92, 0
Centroid of the clusters using TMPLS
Centroid of the clusters using Talairach template
Linear
Nonlinear
Linear
Nonlinear
5.71 5.30 6.64 6.59 4.16 3.35 5.05 6.92
5.65 5.35 5.38 6.58 4.37 4.07 5.22 7.25
7.87 6.77 8.83 8.66 5.94 4.59 5.70 9.17
8.07 7.04 7.58 8.76 6.03 5.27 9.42 6.54
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
415
Fig. 5. Discrepancy between the outputs of MNI2TAL and the Talairach coordinates obtained by the TMPLS approach. The color code represents the discrepancy in the z directory while the arrows show the direction of the discrepancy in the x and y direction. The length of each arrow is proportion to the magnitude of the discrepancy. Note that the scale legend is for the length of the arrow, which is not related to the voxel size.
the MNI have been adopted by the ICBM as an international standard. However, since the Talairach coordinate system is the standard reference system used by the neuroimaging community, it becomes a common practice to report the findings in terms of Talairach coordinates even when the analyzed imaging data have been spatially transformed into different brain templates. As we have demonstrated in this paper, the mapping of a coordinate of the MNI space, based on the ICBM152 template, to the Talairach space generates distinct coordinates for different subjects. However, by computing the centroid of cluster we derived a representative of the coordinate in the Talairach space for a given MNI coordinate and also provide an estimate of uncertainty. Finally, the output of the mni2tal script was compared against the
derived Talairach coordinates, with large discrepancy found in the inferior regions, superior frontal, and occipital regions. By using the mean distance between the cluster centroid and the points within the cluster in the Talairach space, a pattern of uncertainty is revealed. On average the uncertainty value is about 4 mm for the TMPLS approach. Lower uncertainty values are observed for the template-matching approach; on average the value is 0.68 and 2.10 mm for linear and nonlinear transformation, respectively. Similar uncertainty patterns were observed when using either linear or nonlinear transformation to the MNI space. The consistent of the pattern is not surprised since the affine transformation is also performed for the nonlinear spatial transformation in SPM99. The similarity of uncertainty
416
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
suggests that the nonlinear components do not contribute much in the transformation. Unrelated to the uncertainly values, the difference of the derived Talairach coordinates using the linear and nonlinear transformation is similar for both TMPLS and template-matching approaches. The mean difference of the Talairach coordinates is about 1 mm. These results indicate that the method of transforming subject data into the MNI space has less impact on the derived Talairach coordinates but has more impact on the spread of the cluster points. This is true for both TMPLS and template-matching approaches. The obtained results likely depend on the particular transformation of an individual brain into the MNI space. However, the similar pattern found in both linear and the nonlinear transformation provides evidence of the generalization of our results (at least for the SPM99 users). We divided the subject pool into halves and performed the same analysis to investigate whether the uncertainty pattern is biased by the subject pool. The same pattern was found in both analyses, suggesting that the results do not dependent on the selection of the subjects. Without any convenient way to obtain the Talairach coordinate from the MNI space, some researchers rely on the mni2tal script to perform the conversion. As the script author warns, the script can only provide an approximate Talairach coordinate for general reference. By comparing the derived Talairach coordinates with those output by the script, we found that the discrepancies across the brain are in the range of 3.0–9.5 mm for the TMPLS and 2.7– 10.8 mm for the template matching. By examining the coordinates, we found that the outputs of the mni2tal script in z coordinates are always smaller than those derived from the subject space. When interpreting the results coming from the mni2tal script, we suggest special attention must be paid to the regions with the MNI z coordinates less than 12 mm, which have at least 5 mm discrepancy. To allow readers to interpret the results more correctly, we suggest that researchers who analyze data from the MNI space not only report the findings in the converted Talairach coordinates but also the MNI coordinates. Alternatively, researchers may consider using a Talairach-compatible template to avoid the conversion problem (Woods et al., 1999). Since the means to compute the Talairach coordinates are completely different for the TMPLS and template matching, it is not surprising to find discrepancy between the two approaches. The discrepancy values are in the range of 0.18–4.7 mm (mean = 2.82 mm). Similar to the mni2tal output but with smaller magnitude, the relative large discrepancies are found in the inferior regions and the superior frontal and occipital regions.
In this paper, we revisit the issue of the uncertainty of converting the MNI data into the Talairach space. Clearly there is no perfect solution to the conversion problem. However, one can derive the average Talairach coordinates for any point in the MNI space through the individual subject space. With the high operational cost associated with this approach, it is not practical to derive the Talairach coordinates for each study. We are in the process of developing a database that stores the MNI to Talairach mapping derived from a large subject pool and its associated uncertainty measurements and for different age groups, which will be made it freely available through the Internet.
Acknowledgments The authors thank Claude Alain, Steve Arnott, Adrian Restagno, and Kelly McDonald for providing the MRI data; and our imaging analysts, Anda Pacurar and Marc Malloy, for identifying the marker positions for all Talairach transformations. The study was supported by a Canadian Institutes for Health Research NET grant to A.R. McIntosh.
References Ashburner, J., Friston, K.J., 1999. Nonlinear spatial normalization using basis functions. Hum. Brain Mapp. 7, 254 – 266. Brett, M., Johnsrde, I.S., Owen, A.M., 2002. The problem of functional localization in the human brain. Nat. Rev., Neurosci, 3, 243 – 249. Cox, R., 1996. AFNI: software for analysis and visualization of functional magnetic resonance neuroimages. Comput. Biomed. Res. 29 (162), 173. Evans, A.C., Collins, D.L., Mills, S.R, Brown, E.D., Kelly, R.L., Peters, T.M., 1993. 3D statistical neuroanatomical models from 305 MRI volumes. Proc. IEEE Nucl. Sci. Symp. Med. Imaging Conf., 1813 – 1817. 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. NeuroImage 2, 89 – 101. Talairach, J., Szikla, G., 1967. Atlas d’anatomie stereotaxique du telencephale: etudes anatomo-radiologiques. Masson and Cie, Paris. Talairach, J., Tournoux, P., 1988. Co-planar stereotaxic atlas of the human brain. Thieme Medical Publishers, New York. Woods, R.P., Dapretto, M., Sicotte, N.L., Toga, A.W., Mazziotta, J.C., 1999. Creation and use of a Talairach-compatible atlas for accurate, automated, nonlinear intersubject registration, and analysis of functional imaging data. Hum. Brain Mapp. 8, 73 – 79.
The Talairach coordinate of a point in the MNI space: how to interpret it Wilkin Chau* and Anthony R. McIntosh The Rotman Research Institute, Baycrest Centre for Geriatric Care, University of Toronto, 3560 Bathurst Street, Toronto, Ontario, Canada M6A 2E1 Received 18 March 2004; revised 2 December 2004; accepted 6 December 2004
To perform group studies using functional imaging data, the individual brain images are usually transformed into a common coordinate space. The two most widely used spaces in the neuroscience community are the Talairach space and the Montreal Neurological Institute (MNI) space. The Talairach coordinate system has become the standard reference for reporting the brain locations in scientific publication, even when the data have been spatially transformed into different brain templates (e.g., MNI space). When expressed in terms of individual subjects, the mapping of a coordinate in MNI space to the Talairach space generates distinct coordinates for different subjects. In this paper, we describe two approaches to derive the Talairach coordinates from the MNI space, which is based on the ICBM152 template from the International Consortium of Brain Mapping. One approach is the Talairach Method of Piecewise Linear Scaling (TMPLS) as implemented in the AFNI software package; and the other is a template-matching approach using the linear transformation in SPM99. The uncertainty measurements of the mapping results are presented. This may allow researchers to better interpret results reporting in the Talairach coordinates obtained from the MNI space. This study also examines the discrepancy between the derived Talairach coordinates and those obtained from the mni2tal script, a tool commonly used by the neuroimaging community. Large discrepancies are found in the inferior regions, superior frontal and occipital regions. D 2005 Elsevier Inc. All rights reserved. Keywords: Talairach space; MNI space; Spatial normalization; Brain mapping
Introduction Functional neuroimaging studies typically rely on group studies to identify the most consistent pattern of activation associated with a given cognitive operation. The neuroanatomical difference
* Corresponding author. Fax: +1 416 785 2862. E-mail address: [email protected] (W. Chau). Available online on ScienceDirect (www.sciencedirect.com). 1053-8119/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.neuroimage.2004.12.007
between individuals make it important to accurately co-register the brain images of multiple subjects. One approach is based on anatomical landmarks, which requires manually identification of the region of interest (ROI) in each individual brain. This is very costly in terms of time and sometimes ROIs may cross functional boundaries. Most researchers use an alternative approach that based on transformation to a particular brain atlas coordinate system. This approach usually involves an automated process to spatially transform the individual brain image into the coordinate space. However, it does not guarantee that an identical point in the space corresponds to the same anatomical feature for all subjects. The two most widely used spaces in the neuroscience community are the Talairach space (Talairach and Szikla, 1967; Talairach and Tournoux, 1988) and the Montreal Neurological Institute (MNI) spaces (Evans et al., 1993). The Talairach space is based on a stereotaxic atlas of the human brain published by Talairach and Tournoux (Talairach and Szikla, 1967; Talairach and Tournoux, 1988). They identified the anatomical features from the atlas and created a coordinate system related to anatomical landmarks. The Talairach coordinate space has its origin defined at the anterior commissure (AC), with x- and y-axes on a horizontal plan and z-axis on a vertical plane. The yaxis is defined by the line connecting the most superior of AC and the most inferior of the posterior commissure (PC); the x-axis is defined by the line that passes through the AC point and orthogonal to the AC-PC line; whereas the z-axis is the line that passes through the interhemispheric fissure and the AC point. Given a 3-D coordinate in the Talairach space, the anatomical labels can be obtained manually through inspection of the atlas. To accommodate the differences between individual brains, Talairach and Tournoux defined a proportional grid system to align an individual brain to the atlas. The procedure involves dividing the brain into 12 regions by using one horizontal plane, passing through the x- and y-axes, and three vertical planes. Two vertical planes are parallel to the x-axis, one passing through the z-axis and another through the PC point, and one is parallel to the y-axis, dividing the left and right hemispheres. Each region is then scaled, separately for each direction, to match the atlas. This piecewise linear scaling method provides a simple means of
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
converting an individual brain to the Talairach space; however, there is no guarantee that the transformed brain would completely match either the shape or the anatomical features of the Talairach atlas. This linear scaling method performs poorly in terms of matching the anatomical landmark locations to the atlas, especially cortical regions, compared to other methods that use nonlinear transformation. The major criticism about the Talairach atlas is that it was created based on the postmortem brain of single subject, which is not a good representation of the neuroanatomy for the general population. To allow better representation of average neuroanatomy, the MNI created an average brain template based on the MRI scans from several hundred individuals. The first template, known as the MNI305, was created in two steps. The first step was to obtain the average brain of 241 brains; each of them had been reoriented and scaled to match a set of manually selected anatomical landmarks to those of Talairach atlas. The second step was to create the MNI305 template by averaging the 305 normal MRI scans, which had been normalized with a linear transformation matrix to match the average brain created in the first step. Later, a template using the 152 brains normalized to MNI305 template was created. The International Consortium of Brain
409
Mapping (ICBM) has adopted the template as an international standard, known as the ICBM152. This template is used by several functional imaging analysis packages, such as SPM99 (Wellcome Department of Cognitive Neurology, London, UK; http://www.fil. ion.ucl.ac.uk/spm) and FSL (Image Analysis Group, FMRIB, Oxford, UK; http://www.fmrib.ox.ac.uk/fsl). The MNI templates are not perfectly matched with the Talairach standard brain due to large differences in brain shape and size between the two templates. Moreover, individual differences also prompted the ICBM to develop a probabilistic atlas of the human brain (Mazziotta et al., 1995). Since the Talairach coordinate system has become the standard reference for reporting the brain locations in scientific publication, researchers often need to report their findings using this system, even though their data were analyzed in different coordinate space, such as the ICBM152. There is no simple way to transform multiple subject data from the MNI space to the Talairach space. It is quite possible that the coordinate location in MNI space of two subjects would map to different points of Talairach space (Fig. 1a). The discrepancy becomes an issue when the data are analyzed in the MNI space but the results are reported using the Talairach space (Brett et al., 2002).
Fig. 1. (a) Mapping from the SPM99 MNI space to the Talairach space. A point in the MNI space (left) is associated with two different locations of two of the subjects (middle). The locations in the brains correspond to two distinct Talairach coordinates. (b) The schematic diagram of three different ways to obtain the Talairach coordinates of a point in the MNI space. (A) Converted from the MNI space using mni2tal script; (B) using TMPLS approach; (C) using the templatematching approach.
410
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
In this study, we investigate different approaches of deriving the Talairach coordinates from the MNI coordinates and report the uncertainty measurements for the approaches. This study focuses on the coordinate mapping approach, in which the transformation of each individual brain to the common space is independent to the underlying anatomy. More specifically, we studied the results of converting the MNI coordinates, obtained from transforming individual brain images to the ICBM152 template with the SPM99 software, to the Talairach coordinates using two different approaches: (1) the Talairach Method of Piecewise Linear Scaling (TMPLS) using the AFNI software (Cox, 1996; http://afni.nimh. nih.gov/afni); and (2) the linear transformation in SPM99 to match a template that matched the Talairach coordinate space. Finally, we compared the derived Talairach coordinates with those obtained from the mni2tal script (http://eeg.sourceforge.net/mridoc/mri_toolbox/mni2tal.html), a tool commonly used to map the MNI coordinates to the Talairach coordinates (Fig. 1b). This study demonstrates the uncertainty of the transformation from the MNI coordinates to the Talairach coordinates for multiple subject data, showing the spatial extent of the distribution of points in the Talairach space converted from a given MNI coordinates using the above two approaches. Our findings provide a guide for interpretation of the Talairach coordinates obtained from the MNI space. It is especially useful for researchers who conduct meta-analysis using data from both MNI space and Talairach space.
Materials and methods Subjects The MRI scans of 56 subjects (23 females) were obtained from several different fMRI experiments in our institute. Subjects were between the ages of 20 and 38 years (mean age 28 years). All subjects had no history of neurological disorders. Informed consent was obtained from each subject after the nature of each study had been explained. MRI data acquisition All MRI data were acquired with a 1.5-T MR scanner (Signa, General Electric Medical Systems, Waukesha, WI). A standard quadrature transmit/receive birdcage head coil was used. Head restraint was provided using a vacuum pillow (Vac Fix, Par Scientific, Inc.). Anatomical T1-weighted MRI was acquired using
three-dimensional fast spoiled gradient echo imaging technique [124 axial slices 1.4 mm thick, field-of-view (FOV) = 22 22 cm, flip angle/TE/TR = 358/6 ms/16 ms, 1 NEX, 256 256 acquisition matrix]. Data processing SPM99 was used to transform the anatomical data into the MNI space, using the T1 template of ICBM152 provided with the software. The software uses the minimization algorithms to determine the transformation parameters that allow the best match the voxel intensity levels of the template. Two types of transformation were considered: linear and nonlinear. The linear transformation has 12 parameters, which define the translation, rotation, scaling, and shearing operations. The nonlinear transformation is based on a set of cosine basis functions (Ashburner and Friston, 1999). The parameters were set to the default for SPM 99 as follows: the number of basis function was set to 7 8 7, the number of iterations was 13, and the medium regularization was used (i.e., with the value of 0.01). A set of grid points, separated by 4 mm in each direction, in the MNI space were defined as the sample of locations of interest to be compared. To derive the Talairach coordinates for each coordinate in the MNI space, the locations corresponding to the MNI coordinates were identified in each individual brain image. These locations were obtained by extracting the transformation matrices from the SPM spatial normalization operation. In SPM99, the spatial normalization process first determines the coefficients for the affine transformation matrix, and those for the basis functions when nonlinear transformation is performed. These matrices define that transformation from source space (the subject’s brain image in our case) for each sample point to the target space (the MNI space). The corresponding location of individual brain image to each MNI coordinate was obtained by extracting the information from the SPM script of bspm_write_sn.m,Q which contains the source location information for each point in the target space. Once the location in the subject brain was determined, transformations were performed to map the location into the Talairach space. The Talairach Method of Piecewise Linear Scaling (TMPLS) has been adapted in AFNI. Another approach is to map the individual brain to a template that conforms to Talairach coordinate space. Both approaches, the TMPLS and template normalization processes, were used in this study. For the TMPLS approach, the anatomical landmarks of the anterior commissure (AC) superior edge, AC posterior margin, posterior commissure (PC), and two mid-sagittal points, as well as the bounding box of
Table 1 Talairach coordinates of the eight sample locations in the SPM99 MNI space MNI coordinates
64, 16, 12 60, 24, 4 56, 60, 8 48, 72, 40 48, 12, 44 8, 56, 4 12, 24, 64 12, 92, 0
MNI2TAL coordinates
63, 16, 10 59, 23, 5 55, 57, 9 48, 72, 33 48, 10, 41 8, 54, 6 12, 26, 58 12, 89, 4
Centroid of the clusters using TMPLS
Centroid of the clusters using Talairach template
Linear
Linear
58, 18, 14 (3.29) 56, 19, 9 (3.73) 51, 58, 7 (3.84) 43, 72, 35 (4.94) 44, 5, 44 (4.57) 8, 52, 11 (4.03) 10, 27, 60 (4.79) 9, 87, 3 (4.90)
Nonlinear 60, 17, 13 (3.35) 58, 21, 7 (4.21) 53, 57, 9 (4.12) 43, 74, 35 (5.05) 44, 7, 44 (4.79) 9, 50, 11 (4.46) 11, 25, 59 (4.72) 10, 89, 6 (5.08)
60, 15, 17 (0.63) 56, 24, 11 (0.65) 52, 56, 2 (0.83) 44, 69, 41 (0.82) 45, 11, 46 (0.49) 6, 53, 11 (0.59) 10, 24, 63 (0.60) 10, 87, 4 (0.35)
The mean distance (in millimeter) between the cluster centroid and the points within the cluster is shown in bracket.
Nonlinear 60, 15, 18 (2.10) 56, 23, 11 (2.07) 53, 56, 3 (2.06) 44, 69, 41 (2.25) 44, 11, 46 (2.01) 7, 52, 11 (1.68) 10, 24, 63 (2.19) 9, 88, 5 (2.20)
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416 Fig. 2. Clusters in the Talairach space obtained by the TMPLS approach. The corresponding Talairach coordinates of 56 subjects for eight locations in the MNI space. (a) Cluster points from a point of MNI space obtained from a linear transformation. Left: The lateral and medial views of the cluster points. Right: The distribution of cluster points of the superior temporal location; red dot: the cluster centroid, green dot: the output of MNI2TAL script. (b) Same as panel a, but for the MNI data obtained from a nonlinear transformation.
411
412 W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
Fig. 3. Same as Fig. 2, but for the template-matching approach.
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
the brain (i.e., dorsal–ventral, anterior–posterior, and lateral extremes), were manually identified using the AFNI software. From these landmarks, the necessary rotation and linear scaling to transform the brain into the Talairach space is performed. For the template-matching approach, a Talairach template was created using the following procedure: first, each of the 56 brains was transformed into the stereotaxic atlas space using AFNI described above; second, a mean volume was created based on the transformed brains obtained in the step one; and third, the mean volume was smoothed with a Gaussian filter (FWHM = 5 mm) to create the Talairach template. Once the template was created, an affine transformation is performed to map each subject brain to the template using SPM99. Instead of deriving the Talairach coordinates through the subject space, a simple MATLAB script, called bmni2tal,Q written
413
by Matthew Brett (MRC Cognition and Brain Sciences Unit, Cambridge, England), can be used to determine the approximate Talairach coordinates for a point in the MNI space. The conversion is based on two linear transformation matrices, one for the regions above the AC point and the other for the regions below. Unlike the TMPLS and template-matching approaches, the bmni2talQ provides a single Talairach coordinate for each point in the MNI space, which is independent of the subject. Uncertainty measurements Due to difference between the subject brains, the conversion from the MNI space to the subject space, and then to the Talairach space is unique for each subject. A coordinate in the MNI space may be mapped to different Talairach coordinates for different
Fig. 4. The uncertainty measurement of the mapping from the MNI coordinates to the Talairach coordinates using the TMPLS approach. The brain slices are in the MNI space, with z coordinate shown below each slice. The color code represents the uncertainty measurement, which is related to the size of the cluster in the Talairach space corresponding the MNI coordinate. Similar pattern is found when using the template-matching approach.
414
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
subjects. With multiple subjects each point in the MNI space is associated with a cluster of points in the Talairach space. To determine the best Talairach coordinate to represent a given point in the MNI space, the centroid of a cluster points in Talairach coordinates associated with the MNI coordinate was calculated. The mean Euclidean distance between the centroid and the points within the cluster was used to quantify the spatial extent of the cluster. By examining the distribution of the mean distance in the brain, the uncertainty of the mapping from the MNI space to the Talairach space of any given region was determined.
Results Distribution of points in the Talairach space To demonstrate the distribution of the points in the Talairach space, eight points in the MNI space were selected to cover the medial and lateral regions of the brain (Table 1). The Talairach coordinates were obtained by the three approaches described above, namely, (1) the TMPLS approach; (2) affine transformation to match a Talairach template; and (3) the mni2tal scripts. The clusters associated with each MNI coordinate for the first two approaches are shown in Figs. 2 and 3. For the TMPLS approach, the spatial extents of these eight clusters are very similar. The range of the mean distance within each cluster is 3.29–4.94 mm (mean = 4.26 mm) when the linear transformation was used to convert the subject space to the MNI space, and 3.35–5.08 mm (mean = 4.48 mm) for the nonlinear transformation. For the template approach, the spatial extents of the clusters are consistently smaller than those by linear scaling. The range of the mean distance within each cluster is 0.35–0.83 mm (mean = 0.64 mm) for the linear conversion, and 1.68–2.77 mm (mean = 2.15 mm) for the nonlinear one. Uncertainty measurement of the MNI to Talairach space transformation The mean distance between the cluster centroid and each point within the cluster, which is the Talairach coordinates for each subject, is used to measure the uncertainty. For the TMPLS approach, there are two regions corresponding to the largest uncertainty, with values above 5 mm. One is located in the superior frontal region while the other is a band running from the superior occipital to the inferior frontal regions (Fig. 4). The distributions of the uncertainty patterns are very similar for both linear and nonlinear transformations from subject space to the MNI space; the difference of these two patterns is always less than 1 mm. The mean value of the uncertainty is 4.06 and 4.31, respectively. The differences of the Talairach coordinates found associated with the two transformations are always less than 3 mm. From the histograms of the uncertainty values (not shown), more points in the MNI space are found to have values smaller than 4 mm for the linear transformation compared to the nonlinear one. The number of points that have values larger than 4 mm is similar for both transformations. For the template-matching approach, no clear difference in pattern is found associated with the linear and nonlinear transformations. For the linear transformation to the MNI space, the uncertainty values are in the range of 0.3–1.16 with mean = 0.68. For the nonlinear case, the uncertainty values are in the range of
1.45–2.88 with mean = 2.10. Even with different uncertainty values, the differences of the Talairach coordinates found associated with these two transformations are always less than 3 mm, which is similar to those of the TMPLS approach. Performance of the MNI2TAL The performance of the mni2tal program was evaluated through computing the Euclidean distance between the Talairach centroids of the clusters obtained above and the mni2tal output. As an example, the centroid of the sample cluster in the superior temporal region and the corresponding Talairach coordinate found by mni2tal can be found in the scatter plots of Figs. 2 and 3, and the discrepancy values are shown in Table 2. The discrepancy between the centroids and the mni2tal output points is shown in the Fig. 5. The discrepancy between the TMPLS and the mni2tal output is in the range of 3.0–9.5 mm (mean = 6.27 mm) for the linear transformation, and 3.8–9.5 mm (mean = 6.54 mm) for the nonlinear transformation; the discrepancy between the templatematching and the mni2tal output is in the range of 2.7–10.6 mm (mean = 6.14 mm) and 3.6–10.8 mm (mean = 6.69 mm) for the linear and nonlinear transformation, respectively. In general the regions with large uncertainty values are most likely associated with large discrepancy values within each approach. Major discrepancy is observed from the superior frontal region and most of the points with z value b12 mm in the MNI space have discrepancy larger than 5 mm (Fig. 5). Similar pattern of discrepancy is found for both linear and nonlinear transformations from subject space to the MNI space. The discrepancy for the template matching is relatively larger compared to the TMPLS, especially for the inferior regions. For the TMPLS approach, the discrepancy difference between the linear and the nonlinear transformations is in the range of 1.3–1.56 mm (mean = 0.28 mm). For the template-matching approach, the discrepancy difference between the linear and the nonlinear transformations is in the range of 1.3–2.08 mm (mean = 0.54 mm).
Discussion The Talairach and Tournoux atlas has become the de facto standard for localization of in neuroimaging data sets, despite being based on a single postmortem brain. Better representation of the averaging neuroanatomy is achieved through image templates from averaging multiple individuals. The templates developed by
Table 2 Discrepancy between the MNI2Tal and the empirical-derived Talairach coordinates MNI coordinates
64, 16, 12 60, 24, 4 56, 60, 8 48, 72, 40 48, 12, 44 8, 56, 4 12, 24, 64 12, 92, 0
Centroid of the clusters using TMPLS
Centroid of the clusters using Talairach template
Linear
Nonlinear
Linear
Nonlinear
5.71 5.30 6.64 6.59 4.16 3.35 5.05 6.92
5.65 5.35 5.38 6.58 4.37 4.07 5.22 7.25
7.87 6.77 8.83 8.66 5.94 4.59 5.70 9.17
8.07 7.04 7.58 8.76 6.03 5.27 9.42 6.54
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
415
Fig. 5. Discrepancy between the outputs of MNI2TAL and the Talairach coordinates obtained by the TMPLS approach. The color code represents the discrepancy in the z directory while the arrows show the direction of the discrepancy in the x and y direction. The length of each arrow is proportion to the magnitude of the discrepancy. Note that the scale legend is for the length of the arrow, which is not related to the voxel size.
the MNI have been adopted by the ICBM as an international standard. However, since the Talairach coordinate system is the standard reference system used by the neuroimaging community, it becomes a common practice to report the findings in terms of Talairach coordinates even when the analyzed imaging data have been spatially transformed into different brain templates. As we have demonstrated in this paper, the mapping of a coordinate of the MNI space, based on the ICBM152 template, to the Talairach space generates distinct coordinates for different subjects. However, by computing the centroid of cluster we derived a representative of the coordinate in the Talairach space for a given MNI coordinate and also provide an estimate of uncertainty. Finally, the output of the mni2tal script was compared against the
derived Talairach coordinates, with large discrepancy found in the inferior regions, superior frontal, and occipital regions. By using the mean distance between the cluster centroid and the points within the cluster in the Talairach space, a pattern of uncertainty is revealed. On average the uncertainty value is about 4 mm for the TMPLS approach. Lower uncertainty values are observed for the template-matching approach; on average the value is 0.68 and 2.10 mm for linear and nonlinear transformation, respectively. Similar uncertainty patterns were observed when using either linear or nonlinear transformation to the MNI space. The consistent of the pattern is not surprised since the affine transformation is also performed for the nonlinear spatial transformation in SPM99. The similarity of uncertainty
416
W. Chau, A.R. McIntosh / NeuroImage 25 (2005) 408–416
suggests that the nonlinear components do not contribute much in the transformation. Unrelated to the uncertainly values, the difference of the derived Talairach coordinates using the linear and nonlinear transformation is similar for both TMPLS and template-matching approaches. The mean difference of the Talairach coordinates is about 1 mm. These results indicate that the method of transforming subject data into the MNI space has less impact on the derived Talairach coordinates but has more impact on the spread of the cluster points. This is true for both TMPLS and template-matching approaches. The obtained results likely depend on the particular transformation of an individual brain into the MNI space. However, the similar pattern found in both linear and the nonlinear transformation provides evidence of the generalization of our results (at least for the SPM99 users). We divided the subject pool into halves and performed the same analysis to investigate whether the uncertainty pattern is biased by the subject pool. The same pattern was found in both analyses, suggesting that the results do not dependent on the selection of the subjects. Without any convenient way to obtain the Talairach coordinate from the MNI space, some researchers rely on the mni2tal script to perform the conversion. As the script author warns, the script can only provide an approximate Talairach coordinate for general reference. By comparing the derived Talairach coordinates with those output by the script, we found that the discrepancies across the brain are in the range of 3.0–9.5 mm for the TMPLS and 2.7– 10.8 mm for the template matching. By examining the coordinates, we found that the outputs of the mni2tal script in z coordinates are always smaller than those derived from the subject space. When interpreting the results coming from the mni2tal script, we suggest special attention must be paid to the regions with the MNI z coordinates less than 12 mm, which have at least 5 mm discrepancy. To allow readers to interpret the results more correctly, we suggest that researchers who analyze data from the MNI space not only report the findings in the converted Talairach coordinates but also the MNI coordinates. Alternatively, researchers may consider using a Talairach-compatible template to avoid the conversion problem (Woods et al., 1999). Since the means to compute the Talairach coordinates are completely different for the TMPLS and template matching, it is not surprising to find discrepancy between the two approaches. The discrepancy values are in the range of 0.18–4.7 mm (mean = 2.82 mm). Similar to the mni2tal output but with smaller magnitude, the relative large discrepancies are found in the inferior regions and the superior frontal and occipital regions.
In this paper, we revisit the issue of the uncertainty of converting the MNI data into the Talairach space. Clearly there is no perfect solution to the conversion problem. However, one can derive the average Talairach coordinates for any point in the MNI space through the individual subject space. With the high operational cost associated with this approach, it is not practical to derive the Talairach coordinates for each study. We are in the process of developing a database that stores the MNI to Talairach mapping derived from a large subject pool and its associated uncertainty measurements and for different age groups, which will be made it freely available through the Internet.
Acknowledgments The authors thank Claude Alain, Steve Arnott, Adrian Restagno, and Kelly McDonald for providing the MRI data; and our imaging analysts, Anda Pacurar and Marc Malloy, for identifying the marker positions for all Talairach transformations. The study was supported by a Canadian Institutes for Health Research NET grant to A.R. McIntosh.
References Ashburner, J., Friston, K.J., 1999. Nonlinear spatial normalization using basis functions. Hum. Brain Mapp. 7, 254 – 266. Brett, M., Johnsrde, I.S., Owen, A.M., 2002. The problem of functional localization in the human brain. Nat. Rev., Neurosci, 3, 243 – 249. Cox, R., 1996. AFNI: software for analysis and visualization of functional magnetic resonance neuroimages. Comput. Biomed. Res. 29 (162), 173. Evans, A.C., Collins, D.L., Mills, S.R, Brown, E.D., Kelly, R.L., Peters, T.M., 1993. 3D statistical neuroanatomical models from 305 MRI volumes. Proc. IEEE Nucl. Sci. Symp. Med. Imaging Conf., 1813 – 1817. 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. NeuroImage 2, 89 – 101. Talairach, J., Szikla, G., 1967. Atlas d’anatomie stereotaxique du telencephale: etudes anatomo-radiologiques. Masson and Cie, Paris. Talairach, J., Tournoux, P., 1988. Co-planar stereotaxic atlas of the human brain. Thieme Medical Publishers, New York. Woods, R.P., Dapretto, M., Sicotte, N.L., Toga, A.W., Mazziotta, J.C., 1999. Creation and use of a Talairach-compatible atlas for accurate, automated, nonlinear intersubject registration, and analysis of functional imaging data. Hum. Brain Mapp. 8, 73 – 79.