Landmark-Based Morphometrics of the Normal Adult Brain Using MRI

NeuroImage 13, 801– 813 (2001) doi:10.1006/nimg.2001.0748, available online at http://www.idealibrary.com on

Landmark-Based Morphometrics of the Normal Adult Brain Using MRI S. L. Free,* P. O’Higgins,† D. D. Maudgil,* I. L. Dryden,‡ L. Lemieux,* D. R. Fish,* and S. D. Shorvon* *Epilepsy Research Group, National Society for Epilepsy MRI unit, Institute of Neurology, Queen Square, London; †Department of Anatomy and Developmental Biology, University College, London; and ‡Department of Statistics, Nottingham University, Nottingham, United Kingdom Received July 18, 2000; published online March 19, 2001

We describe the application of statistical shape analysis to homologous landmarks on the cortical surface of the adult human brain. Statistical shape analysis has a sound theoretical basis. Landmarks are identified on the surface of a 3-D reconstruction of the segmented cortical surface from magnetic resonance image (MRI) data. Using publicly available software (morphologika) the location and size dependence of the landmarks are removed and the differences in landmark distribution across subjects are analysed using principal component analysis. These differences, representing shape differences between subjects, can be visually assessed using wireframe models and transformation grids. The MRI data of 58 adult brains (27 female and 15 left handed) were examined. Shape differences in the whole brain are described which concern the relative orientation of frontal lobe sulci. Analysis of all 116 hemispheres revealed a statistically significant difference (P < 0.001) between left and right hemispheres. This finding was significant for right- but not left-handed subjects alone. No other significant age, gender, handedness, or brain-size correlations with shape differences were found. © 2001 Academic Press

INTRODUCTION Structural descriptors of the brain have generally been of two types. First are the volumes, areas, and lengths of particular anatomical structures, where much emphasis is placed on the accurate and reproducible definition of the boundaries of those structures and their isolation from neighbours (Blatter et al., 1995; Buchanan et al., 1998; Davatzikos and Resnick, 1998; Hutsler et al., 1998; Jack et al., 1997; Kennedy et al., 1998; Kulynych et al., 1996; Raz et al., 1997). Second, the analysis of groups of brains registered together in a common space where the entire brain or large regions of it are compared, voxel by voxel, with statistical techniques, between many subjects (Abell et al., 1999; Ashburner et al., 1998; Gaser et al., 1999; Woermann et al., 1999). These latter rarely consider

the specific anatomical aspects of substructures as they encompass the whole brain in one analysis. The limitations of the first technique are the generally timeconsuming and subjective nature of detailed anatomical definitions, leading to the consideration of only isolated structures and rarely allowing features of global shape differences to be assessed. The limitations of the second are that, as no anatomical correspondence is employed, differential changes in specific substructures of the brain are overlooked, limiting the biological conclusions that can be drawn. One set of analytical approaches that avoids some of these limitations is geometric morphometrics or statistical shape analysis. This term has come to be applied to approaches pioneered in the statistics community and subsequently exploited in the study of evolutionary biology (e.g., O’Higgins, 2000). In these analyses landmarks (homologous points) are chosen over the structure of interest and a statistical analysis examines variations in landmark configurations. Statistical differences in the configurations can be examined and visualised and may be used for a biological interpretation of shape differences amongst subjects (Bookstein, 1989; O’Higgins and Jones, 1998). Statistical shape analysis is dependent upon and gains its power from the accurate identification of homologous landmarks. Most morphometric studies, to date, have used landmarks collected on bony structures where equivalence can often be clearly determined (O’Higgins and Jones, 1998). Few studies have used landmarks on the brain. The difficulties in, and our approach to, identifying landmarks on the brain cortical surface (defined by MRI) have been discussed previously (Maudgil et al., 1998). Applications of geometric morphometrics have come primarily from the work of evolutionary biologists but some applications to medical imaging have been studied (Bookstein, 1997). DeQuardo et al. considered landmarks around the corpus callosum and brain stem in a single mid-sagittal MRI image (DeQuardo et al., 1999). A transformation grid deformed with thin plate splines demonstrated differences in the corpus callosum be-

801

1053-8119/01 $35.00 Copyright © 2001 by Academic Press All rights of reproduction in any form reserved.

802

FREE ET AL.

tween normal controls and patients with schizophrenia. Buckley et al. also assessed patients with schizophrenia, considering landmarks obtained from the ridge curves defining the ventricular system in MRI data (Buckley et al., 1999). Again statistically significant differences between patients and controls were identified and demonstrated using transformation grids. Le Goualher et al. recently considered variation in the median surface of the central sulcus, as defined by a parametric mesh (Le Goualher et al., 2000). They used a principal component analysis to compare surfaces across a group of subjects comprising mono and dizygotic twins and unrelated subjects. Surfaces had first been matched with a Procrustes fit but with no specific anatomical correspondence determined. No obvious pattern of difference between subjects was observed except that monozygotic twins were more significantly similar to each other than dizygotic twins and unrelated subjects. In this study we have applied the techniques of statistical shape analysis to the study of the normal adult human brain using landmarks distributed over the cortical surface to assess shape differences in the disposition of sulci. In addition, the relation of landmark variations to other characteristics of the subjects (age, gender, size, and handedness) are considered. MATERIALS AND METHODS Subjects and Data Fifty-eight normal control subjects were assessed, none of whom reported a history of neurological or psychiatric illness. Handedness was determined by a 13-point questionnaire about hand usage in common situations (Chapman and Chapman, 1987). Subjects were identified as right handers or non-right-handers. Twenty-seven subjects were female (19 were righthanded, 8 non-right-handed). Thirty-one subjects were male (24 were right-handed, 7 non-right-handed). The age range of subjects was 16 to 59 years (mean 33 years, median 31 years). Non-right-handers were gender and age-matched (mean difference less than 2 years) with right-handers. Independent t tests found no significant differences in age between males and females, nor between right- and non-right-handers. All subjects were scanned on a 1.5T MRI scanner (GE Medical Systems, Milwaukee, WI). Thirty-one subjects had a T1-weighted fast spoiled gradient echo (FSPGR) volume sequence [TR/TE 35/5 ms, 35° flip angle, 256 ⫻ 128 matrix, 24 ⫻ 24-cm field of view] on one scanner and 27 subjects had a T1-weighted inversion recovery prepared FSPGR volume sequence on another scanner [TR/TE/TI 17.4/4.2/450 ms, 20° flip angle, 256 ⫻ 192 matrix, 24 ⫻ 18-cm field of view]. All data sets contained 124 slices of 1.5-mm thickness, acquired in the coronal plane through the entire brain.

A scan origin label for each subject was included as an extra factor in the subsequent statistical analysis. Processing and Visualisation Data were transferred to an independent workstation (Allegro; ISG Technologies, Toronto, Canada), where they were automatically transformed to cubic voxels, and a standard contrast range was set to optimise, visually, the contrast between grey and white matter. This also effectively provided a relatively constant contrast between grey matter and cerebrospinal fluid at the pial surface. A semiautomated segmentation of the cortical surface was applied to each data set. A seed point is placed within the brain on an MRI slice. The Allegro software grows a region of interest about the seed which is limited by a visually determined intensity threshold, i.e., at the pial surface of the grey matter. The resulting region of interest (ROI) was manually edited to separate the left and right hemispheres from each other and from the cerebellum and to remove any overlying meninges. Cerebral hemispheres were separated from the brain stem and other basal structures following a standardized anatomical protocol (Sisodiya et al., 1995). Reliability of segmentation has been estimated previously by repeat segmentations and estimation of cerebral hemisphere volume. The interrater correlation coefficient for eight studies analyzed by two raters was 0.92 for each hemisphere (Sisodiya, 1996). The regions of interest describing the left and right hemispheres of each subject were transferred to a Sun workstation (Sun Microsystems, Mountain View, CA) and converted to an Analyze 7.5 image format (Mayo Foundation, MN) with a slice thickness defined by the cubic format of the Allegro. This enabled the ROI, now volumes of interest, to be reconstructed in three dimensions to allow the identification of landmarks on the cortical surface. Standard software enabling a volume reconstruction uses a lighting model to give the impression of the 3-D surface on the flat computer screen. To avoid the position-dependent effects of a lighting model we used in-house software to calculate the local mean curvature on the brain surface according to a published algorithm (Florack, 1993). The resulting 3-D reconstruction, without the use of a lighting model is white (or positive) for voxels on the convex parts of the brain surface (gyri) and black (or negative) for voxels on the concave parts of the brain surface (sulci). The 3-D reconstruction can be rotated in the three planes of the data and 3-D coordinates on the surface can be identified when the cursor is clicked on the reconstructed surface. We identified 12 cortical surface landmarks on each hemisphere, that is 24 landmarks per subject, (see Table 1 and Fig. 1). The choice of landmarks and criteria used to determine their suitability as homologous

MORPHOMETRICS OF THE ADULT BRAIN USING MRI

TABLE 1 Landmarks in a Brain Hemisphere Used for Morphometric Analysis 1 2 3 4 5 6 7 8 9 10 11 12

Termination of cingulate sulcus at the midline Superior rostral sulcus intersection with anterior cingulate Parieto-occipital sulcus to the midline Parieto-occipital sulcus intersection with the calcarine Central sulcus to the midline Precentral sulcus to the midline Precentral sulcus intersection with the superior frontal sulcus Central sulcus intersection with the Sylvian Fissure Postcentral sulcus intersection with the Sylvian Fissure Precentral sulcus intersection with the Sylvian Fissure Precentral sulcus intersection with the inferior frontal sulcus Preoccipital notch

landmarks has been discussed previously (Maudgil et al., 1998). These landmarks were chosen for their reliability of identification in a subset of normal control brains, both in anatomical terms and with respect to the limitations of the MRI data and the segmentation process. They principally represent the intersection of named major sulci or their terminations. Four landmarks were on the medial surface and the rest predominately on the anterior/lateral surface. Sulci on the basal surface tend to be compressed by the weight of the superior brain tissue and gyri abut very closely. This meant that the sulcal folds on the basal surface could rarely be clearly seen in the segmented data. MR susceptibility artefact in the basal temporal region also reduced segmentation quality in this part of the brain. Gyri on the occipital/posterior part of the brain also tend to abut very closely which impinges on the clarity of the segmented data. Additionally the anatomical labelling of occipital sulci is problematic as the minor sulci in this area tend to be more variable between subjects (Duvernoy, 1991). Sulci in the temporal lobe were also excluded. These sulci can be clearly seen along most of their length but their sulcal terminations are difficult to observe and they do not have any significant sulcal intersections. This meant that determining a consistent landmark on these sulci was much more prone to error. Ultimately we preferred to choose a limited number of landmarks which we could confidently identify than a larger number with greater variability. The chosen landmarks were identified on the surface of the cortex as either the termination of a sulcus or the intersection of two sulci. Variation in gyral abuttal means that segmentation to the fundus of a sulcus was variable between subjects. Thus to avoid artefactual differences in landmark position within the depth of the sulcus, the landmarks were always chosen at the lip of the sulcus. For the termination of sulci, the anterior edge of the sulcus was chosen (in the case of the parieto-occipital sulcus, which was the only chosen sulcus which commonly bifurcates, the anterior edge of

803

the anterior branch was always chosen). For terminations at the midline it was often necessary to identify the coordinates in one plane, for example, on the medial surface and then rotate the brain to a view perpendicular to the medial surface to check the ordinate in that plane. For landmarks defined at the intersection of sulci the anterior superior aspect of the intersection was chosen as the landmark. If sulci did not intersect in a particular subject (for example, the inferior termination of the central sulcus does not always communicate with the Sylvian fissure) it was necessary to extrapolate visually to label a landmark where intersection would take place if the sulcus had continued. Inspection of our previously reported repeatability measures for landmark identification (Maudgil et al., 1998), led to improvements in our anatomical labelling protocol which was used in the present study. The order of landmark identification was consistent across all subjects and proceeded according to the order in Table 1. Surface views at standard orientations were also employed for each landmark, for example, at 20° from the upright to view the intersection of the superior frontal and precentral sulci. This standardization is limited, however, as brains were not specifically oriented during data acquisition. The termination of the central sulcus to the midline gave rise to the greatest interrater variation in our initial study—this was invariably because it was confused with the precentral sulcus. Kido et al. found the termination of the cingulate always posterior to the central sulcus in a study of postmortem brains (Kido et al., 1980). Our revised landmarking protocol ensures that the termination of the cingulate, which is an easier landmark to identify, is labeled first, which gives more confidence in subsequently locating the central sulcus. Nine randomly chosen subjects were measured twice, throughout the data collection process, to obtain repeatability information for the landmark identification and the entire morphometric analysis. Morphometric Analysis Morphometric analysis was carried out on all 58 subjects with 24 landmarks for each brain to assess shape differences of the whole cerebrum between subjects. In addition the landmark data for the right hemisphere was reflected across the midsagittal plane yielding 116 landmark configurations of 12 landmarks each. These were examined to determine shape differences between the hemispheres. Landmark coordinates were collected in a data file with the subject information and analyzed morphometrically in morphologika [http://evolution.anat. ucl.ac.uk/morph/]. The steps in our morphometric analysis have been outlined elsewhere (O’Higgins and Jones, 1998) with further mathematical detail in Dry-

804

FREE ET AL.

FIG. 1. Three views of the left hemisphere of one subject indicating the approximate locations of the cortical landmarks, generally at the anterior/superior edge of the sulcal intersection. The data are volume renderings of the mean curvature of the cortical surface, displayed with no lighting model. Top, the medial surface. Middle, the superior surface. Note in this subject the discontinuous nature of the precentral sulcus. The landmark at the precentral sulcus to the midline is chosen as the sulcus, which reaches the midline and is the most direct continuation of the precentral sulcus from further inferior. Bottom, the lateral surface. Note in this subject that the central sulcus does not reach the Sylvian fissure. The landmark is chosen at a point where the sulcus would reach the fissure if it was extrapolated.

den and Mardia (1998). To estimate the relative variation among the sampled landmark configurations the data sets need to be registered to each other to remove dependence upon size and location and orientation in space. Although size may be of interest in any biological interpretation we want to separate it from the shape variation. Size, or centroid size, is defined as the square root of the sum of the squared Euclidean distances from each landmark to the centroid of the landmarks. We have used a generalized Procrustes analysis to register different landmark configurations. This fit minimises the sum of the squared distances between

k landmarks in m dimensions from n landmark configurations. Centroid size was used as a surrogate for brain size when considering the correlation of brain size with shape features in the subsequent statistical analysis. The initial landmark configurations occupy a space of k ⫻ m dimensions. The Procrustes transformed coordinates occupy a space with reduced dimensions—a so-called shape space first described by Kendall (1984). As a simple example, for a configuration with k ⫽ 3 landmarks in an m ⫽ 2 dimensional plane (triangle), this shape space is the surface of a sphere with each triangle represented by a single point on that surface. All possible shapes defined by three landmarks are represented on that surface. For objects defined by more than three landmarks the shape space is more complex, with (km ⫺ m ⫺ 1 ⫺ (m ⫺ 1)/2) dimensions. The multiple dimensions and curvature of the shape space make it a difficult one in which to analyze object variability. One solution is to extract the data points to a tangent plane to the shape space in which standard multivariate statistical techniques can be applied. This approximation means that the ensuing statistical analysis will be dependent upon the projection method used. In our study we have used the full Procrustes tangent space projection as proposed by Dryden and Mardia (1993). The validity of the approximation in our data was tested. The tangent plane is defined at a pole in the shape space which is chosen to be the mean of the Procrustes fitted data. In the tangent space the variability between landmark configurations, that is the variability between subjects, is analyzed with a principal components analysis of the tangent coordinates. A principal component (PC) score is generated for each subject for each PC using the eigenvectors of the PC analysis. While principal component analysis can be used to visually demonstrate shape differences in the landmark configurations it does not indicate whether these shape differences are statistically significant. The difference between groups within the data set can be assessed by Goodall’s F test, (Goodall and Mardia, 1991) and by permutation analysis, both using the Procrustes fitted data (Good, 1993). For the permutation test, the Full Procrustes distance (Euclidean distance between Procrustes fitted shapes) is calculated between the groups of interest and then between groups of the same size randomly generated from the entire data set. The ranking of the distance of interest within all calculated distances (1000 iterations) is used to ascribe a statistical significance to the distance/difference between the groups of interest. The relative contribution of individual landmarks to the shape differences can be assessed by calculating the mean left and mean right hemisphere landmark configurations (after a separate Procrustes fit for each hemisphere) and then performing a Procrustes fit between the two

MORPHOMETRICS OF THE ADULT BRAIN USING MRI

FIG. 2. Principal component analysis of data from 24 cortical landmarks in 58 subjects, after Procrutes fit and tangent plane projection. Each point represents one subject, defined by their score on the first two principal components, i.e., the axes of greatest difference in the data set. Male right-handed subjects (circles); female right-handed subjects (crosses); male left-handed subjects (stars); female left-handed subjects (triangles).

mean configurations (to generate the mean of the means). The Euclidean distance of each landmark from the hemisphere mean to the mean mean can be calculated. The greater this distance the more variation between subjects in the relative position of that landmark. Visual Interpretation

805

of smaller ones of similar importance. None of the first five principal components showed any significant correlation with age, gender, brain size (centroid size), or handedness. Figure 2 is a graph of principal component 2 (PC2) against principal component 1 (PC1) for all subjects. Male and female, right- and left-handed subjects are indicated by different symbols. Figure 3 shows PC2 against PC1 for all subjects including repeat measures (n ⫽ 58, plus 9 repeats). In this plot subjects measured more than once are shown with different markers. This indicates that repeat measurements of the same specimen are generally close to each other in the parametric space. For repeat measures, the mean Euclidean distance of landmarks from the mean landmark position (mean of repeats) ranges from 0.25 to 1.55 mm. This is less than in our earlier study, when the mean of such measures, across all landmarks was 3.7 mm for intrarater variability and 6.0 mm for interrater variability (Maudgil et al., 1998). This reflects the improvements to the labeling protocol after our initial study. Interrater variability was not reassessed in this present study. MRI scans were acquired on two different scanners in this study, with slightly different scan sequences used on each scanner. There was no permuted difference in landmark distributions nor any group separability observed in the first 10 principal components of the PCA on the basis of sequence type. With respect to the validity of assumptions in using the tangent projection, all scores of the first two principal components for our whole brain data lie at values less than 0.12 from the mean (the origin). In Fig. 4 the shape differences corresponding to changes along the first principal component axis are shown (for all subjects without repeat measures). For

To visualize the shape variability associated with each principal component an inverse procedure generates landmark configurations from PC scores. Tangent coordinates can be defined for specific PC scores and the inverse tangent projection transformation enables the estimation of landmark coordinates in the original configuration space. The landmarks can be connected by lines to define a wireframe model which aids visual interpretation. Additionally the variation in landmark configurations due to variation in PC score can be used to deform a square grid placed in the image space. A thin plate spline provides a good model for smooth deformations of the grid (Bookstein, 1989). RESULTS A principal component analysis of the 58 brains (24 landmarks each) indicates that 40% of all shape variance is explained by the first five PC scores (PC1 10.3%; PC2 9.5%; PC3 7.1%; PC4 7.0%; PC5 6.1%). These values indicate that the data are not strongly determined, that is, there is no single dominant aspect of shape variability between these brains but a number

FIG. 3. Principal component analysis of data from 24 cortical landmarks in 58 subjects, with 9 additional points for subjects in whom landmark data was collected twice. All subjects marked with circles except 9 subjects with other markers representing repeat measurement data.

806

FREE ET AL.

the negative extreme of PC1 the superior midline landmarks in the right hemisphere are posterior to those in the left hemisphere; for positive PC1 these landmarks are anterior to those in the left hemisphere (see wireframe shapes in Figs. 4a and 4b; transformation grid at level of superior midline landmarks in Fig. 4c). The landmark at the junction of the pre central sulcus and the Sylvian fissure (landmark 10, see Table 1) is relatively anterior for negative PC1 and has moved more posteriorly for positive PC1. In Figs. 4d and 4e, the lateral view of the right hemisphere shows that for positive PC1 (Fig. 4e), when the central and pre central sulci are more vertical the distance from superior midline to preoccipital notch (landmarks 5 to 12) is greater than for negative PC1 (Fig. 4d), this is similar for the left hemisphere. Taken together, the variations along PC1 differentiate between brains with more vertical orientation of frontal lobe features and those with more horizontal orientation of features, i.e., spread more in the anteroposterior plane. For the analysis of the 116 hemispheres (12 landmarks each), 46% of all shape variance is explained by the first 5 PC scores (PC1 14.1%; PC2 10.1%; PC3 9.5%; PC4 6.6%; PC5 6.1%). Again the data are not strongly determined. In Fig. 5 scores for PC5 against PC2 are shown for the analysis of all 116 hemispheres with left and right hemispheres marked by different symbols. A separation of left and right hemispheres in the PC space is shown (left hemispheres tending to positive PC5 and negative PC2 and right hemispheres to the opposite). The statistical significance of this difference is confirmed by a permutation test of left against right hemispheres with a significance of P ⬍ 0.001 and by a Goodall’s F test statistic of 3.915 and P ⬍ 0.000. This significant difference is maintained for right handed subjects alone but not for left handed subjects alone; however, the number of left handers was small. The shape differences demonstrated by the PC analysis of the hemispheres do not correlate with age, gender or brain size. Centroid size for the right and left hemispheres within subjects were significantly correlated with an r 2 of 0.79. The shape difference associated with the separation of left and right hemispheres is illustrated in Fig. 6 in three views of the wireframe shapes. The most obvious difference is in the position of the pre-occipital notch which is more posterior and medial in left more often than right hemispheres. The calculation of Euclidean distances for individual landmarks from the mean landmark configuration indicated that landmarks 12, 4, and 2 are the most variable in both hemispheres (that is the preoccipital notch, parieto-occipital sulcus intersection with the calcarine sulcus and the superior rostral sulcus intersecting with the cingulate sulcus, respectively). Landmarks 5 and 8 contribute least to the variation (that is the central sulcus at the midline

and the central sulcus intersecting with the Sylvian fissure, respectively). DISCUSSION We have used geometric morphometric analysis of 24 landmarks in 3 dimensions to consider the shape of the human cortex. Other studies have used these same or similar techniques to assess the corpus callosum in cross-section, the ventricular system in 3-D and the ribbon of the central sulcus (Buckley et al., 1999; DeQuardo et al., 1999; LeGoualher et al., 2000; Manceaux-Demiau et al., 1998). We have used a previously devised cerebral landmarking system, with good repeatability, and collected data in 58 adults with no neurological or psychiatric disorder. Morphometric analysis of the resulting data identified significant differences between the left and right hemispheres of the brain across all subjects. These shape differences were neither age nor gender dependent. There was no size/ shape correlation. Shape differences were associated with handedness but the small number of left handed subjects limits the conclusions about this relationship. Our subjects were scanned as part of a study into cortical malformation associated with epilepsy. Thus no specific effort was made to recruit nonright handers which would have improved the statistical comparison of subjects for this factor. Our definition of handedness relied only on self-reported preference according to a written questionnaire. While this has been validated and is in use in our centre in a clinical setting this does not provide as rigorous a definition as that obtained from manual testing. In the literature more detailed studies have therefore been able to classify subjects as right handed, left handed and mixed handers (Amunts et al., 2000). Our data do not allow this sub-division and presumably mixed handers are split between our two groups. As data were obtained over some time in a clinical setting two different MR sequences were used. Permutation analysis and the PCA labeled for scan origin did not reveal any significant differences between data sets acquired from different scans. This suggests that either the different sequences did not affect the landmark identification or that other sources of variability including cross-subject anatomical variability contributed more to the landmark variation. There was good visualization of many sulci and gyri in the frontal and lateral regions on the postprocessed MRI but not on the basal and posterior regions of the brain. Poor visualization is likely due to a number of factors which include susceptibility artefacts and operator dependence in the segmentation process. However even seeing the sulci clearly does not necessarily enable straightforward anatomical labeling. During the development of the landmarking protocol it was clear that repeatability of the landmark coordinates was much more dependent on the correct identification of

MORPHOMETRICS OF THE ADULT BRAIN USING MRI

the sulci rather than the correct identification of the landmark placement on the sulci. Thus repeated coordinate values were either (most commonly) almost exactly the same or significantly different (always when one sulcus was mistaken for another). This led to the anatomical labelling protocol being refined to enhance the ability to correctly label sulci as detailed in the methods. One significant limitation of our study was the small numbers of landmarks chosen. We chose a small number of landmarks about which we had confidence rather than a larger number for which labelling would have been less rigorous and homology less certain. We think therefore that the analysis should be robust, that is we would obtain similar results in another group of subjects, but will be limited in the scope of its description. It can only give us information about the relations between chosen landmarks. Thus important areas of the brain were excluded from the analysis. The reason for a limited choice usually depended upon the confidence of the anatomical labeling. Recent advances in automatic sulcal atlas labeling would offer a route to improving this part of the protocol (Royackkers et al., 1999; Sandor et al., 1997). A further limitation is the restriction of landmarks to the cortical surface. Recent studies have identified significant intersubject differences related to the cortical depth of several sulci (Lohmann et al., 1999; Amunts et al., 2000). Sulcal depth provides additional information about the developmental processes of the brain and is likely to be significant in disease processes. Our sulcal labeling protocol using 3-D reconstructions of cortical surface inevitably made assessment of cortical depth more difficult. A combination of 2-D and 3-D data would allow easier identification of points throughout the whole cerebrum, including deep cortical structures. This would aid interpretation of the biological significance of shape changes. With regard to the validity of our choice of the tangent projection in the morphometric analysis, Rohlf has demonstrated that if the data sets are close in the shape space (Dryden and Mardia suggest data within a full Procrustes distance of 0.2 from the mean shape) all tangent projections should give similar answers (Rohlf et al., 1999). As our data (for PC1 of the whole brain analysis) all lie at values less than 0.12 from the mean, our data are sufficiently confined within shape space that the assumptions necessary to use tangent projections are justified (Dryden and Mardia, 1998). By collecting data on each hemisphere separately we were able to explicitly compare left against right hemispheres. The most significant finding of our analysis in normal human brains concerns the symmetry of the hemispheres. Functional asymmetries in brain function in normal controls are well known and have generated interest in identifying their structural correlates. In two studies Ide and colleagues have assessed

807

the bifurcation patterns in the Sylvian fissure and patterns of other sulci in 40 post mortem brains (Ide et al., 1996, 1999). Specific patterns could be identified in some sulci, some correlating with gender and hemisphere. However, a specific pattern in one hemisphere could not predict the pattern on the other hemisphere, suggesting independent development of fissures in the hemispheres. Our findings add support to this by showing that in any one subject the right hemisphere is usually more comparable to the right hemisphere of other subjects than to the left hemisphere of the same subject (see Fig. 5). Asymmetries in the extent and shape of the Sylvian fissure in post-mortem brains have also been quantified by two studies, one using manual measurements on the surface (Witelson and Kigar, 1992) and one using image analysis with parametric meshes of the full sulcal extent (Thompson et al., 1998). Unfortunately we did not have landmarks at the posterior end of the Sylvian fissure where these differences were predominant and this is related to the fact that other studies found significant variation in the location of posterior terminal landmarks. That is, we have avoided regions of ambiguous anatomical structure because we could not find reliable landmarks. This is a limitation in that some regions of greater difference will not be assessed by this method. However, the landmarks chosen in this study all refer to major sulci and changes in these may reflect more significant differences in brain structure rather than a more common variation in mainly minor sulci. Our interpretation must be guided by the fact that regions of significant variation may have been excluded and we can offer only a partial explanation of brain differences. Several studies have shown symmetry in whole hemisphere volumes (Sisodiya et al., 1995) or surface area (Tramo et al., 1995) with which our centroid size data is in agreement. However, substructure volume or surface area asymmetries have been identified for the planum temporale (Kulynych et al., 1995; Rumsey et al., 1997), planum parietale (Jancke et al., 1994), primary auditory cortex (Penhune et al., 1996), superior temporal gyrus (Kulynych et al., 1996) and postcentral and cingulate gyri (Hutsler et al., 1998), and numerous other named gyri (Tramo et al., 1995). An alteration in the relative placement or size of substructures (i.e., cortical shape) must be necessary to accommodate such substructure volume or area asymmetries within whole hemisphere symmetry. Planum temporale asymmetry has recently been repostulated as a orientation difference only which gives rise to area or volume differences because of changes in landmark position (Westbury et al., 1999). Our study can not be directly related to all these findings since we did not obtain landmarks in all areas relevant to these previous studies. However, we have shown additional variation in the rel-

808

FREE ET AL.

FIG. 4. Wireframe models of shape differences estimated by the principal component analysis of 24 landmarks in 58 subjects (from Fig. 2). The landmarks are at the vertices of the wireframe and are numbered as they appear in the table. a and b are wireframes of the landmark distributions at either end of the first principal component axis (a at negative PC1 and b at positive PC1). The wireframes are viewed from vertex. c shows a transformation grid overlying the wireframe model. The grid represents the shape difference which occurs when transforming from the most negative PC1 value to the most positive PC1 value. The grid is positioned at the level of the superior midline landmarks. d and e are the same as a and b, but with the wireframe viewed from the lateral side.

ative placement of sulci (through the surrogate of landmarks along those sulci) with consistent differences between the left and right hemispheres. The most obvious aspect of this difference is in the relative position of the preoccipital notch and the intersection of the calcarine and parieto-occipital sulci. Zilles et al. identified the mean left central sulcus as being more posterior with respect to the right in MRI planar images in 28 male right handed subjects (Zilles et al., 1996). This agrees with our finding in a propor-

tion of the brains studied (Fig. 4). However, our analysis of which landmarks contributed most to the differences between the mean left and mean right hemispheres found the least variation in the two landmarks at either end of the central sulcus. This suggests that on average the position of the central sulcus in a group of subjects does not vary greatly between the hemispheres but in any one subject there may be observable variation. White et al. did not detect any hemispheric asymmetry in the central sulcus from

MORPHOMETRICS OF THE ADULT BRAIN USING MRI

FIG. 5. Principal component analysis of data from 12 cortical landmarks in 116 cerebral hemispheres in 58 subjects, after Procrutes fit and tangent plane projection. Each point represents the landmark configuration from one hemisphere, defined by its score on the 2nd and 5th principal components. These principal components give the greatest discrimination between left and right hemispheres. Left hemispheres are circles, right hemispheres are crosses. Hemispheres from any one subject are not coincident in the PC space.

measures on 67 postmortem brains, except a nonsignificant tendency to a longer fundus in the left hemisphere (White et al., 1997b). This is consistent with our identification of a more horizontal rather than vertical orientation in some subjects (Fig. 4). They had no handedness information for further correlations. The two studies of Amunts et al. contradict White in that more asymmetry of central sulcus depth was found in male control subjects (Amunts et al., 1996, 2000). The left central sulcus was deeper than the right in right handed males and vice versa for left handers. There was no asymmetry for mixed handers. However, no asymmetries were seen in females whatever their handedness. A study by Zilles found that the frontal pole protrudes more in the right than left hemisphere of right-handed males but is not so asymmetric in left handers (Zilles, 1996). We found a significant asymmetry between hemispheres in right handers which was not maintained in left handers (i.e., they were less asymmetric) and the shape difference associated with this differs from that observed by Zilles group. However, sample size may confound this— our study has more right handers and fewer left handers than the Zilles study. The difficulties of obtaining sufficient left handed subjects has meant that left handers are often specifically excluded from studies in the literature in case they introduce confounding factors which limits the number of studies for comparison. Thompson et al. considered parametric meshes from the manual delineation of the medial axis of several sulci in six postmortem brains (Thompson et al., 1996).

809

The occipital sulci (parieto-occipital and calcarine) varied more in the superior/inferior direction whereas paralimbic sulci varied more in the anterior/posterior direction. We do not have paralimbic sulcal landmarks for comparison and found the greatest variation in the landmark at the intersection of the parieto-occipital sulcus and the calcarine sulcus in the anterior-posterior not the superior/inferior direction. This may reflect either the smaller numbers in the Thompson study or their use of whole sulci rather than points. Le Goualher et al. used principal component analysis to identify greater similarity in the shape of the central sulcus within monozygotic than dizygotic twin pairs (Le Goualher et al., 2000). This similarity was stronger in the left hemisphere. How this may be correlated with the findings of Amunts (above) is not clear as no handedness information was provided by Le Goualher et al. However, it seems feasible that a deeper sulcus may vary less in overall shape, due to, if nothing else, the mechanical constraints of more surrounding tissue at greater depths. Specific anatomical correspondence was not assured in the Le Goualher et al. study, but this may only be possible or meaningful along a sulcus by consideration of cytoarchitectonic boundaries which are not seen on MR data (White et al., 1997a). Another study of sulcal variation in twins by Lohmann et al. considered only monozygotic twins. After nonlinear deformations to remove some, but not all of, the gross brain shape differences, the sulcal depth and relative neighbourhood matching of sulci on the lateral surface were assessed. A greater similarity in twin pairs than unrelated pairs was found. There was also greater similarity in the posterior than anterior sulci in all subjects and this was more marked for the left hemisphere. The similarity within twin pairs also increased with increasing sulcal depth. Handedness was not reported. We can not compare with our study directly as we had only one landmark in the posterior lateral region and Lohmann’s study did not include the medial surface. Greater similarity in posterior rather than anterior sulci is in contrast to atlas studies which found greater variability in the occipital sulci (Duvernoy, 1991). In Lohmann’s study no explicit anatomical correspondence is defined so it is probable that sulcal ‘matches’ are made between different sulci in the different subjects. This approach has utility in assessing cases where there is a presumptive similarity ie. monozygotic twins. However, the matching of non-homologous sulci would make interpretation of other shape issues more problematic. If the central sulcus in one subject is matched with the precentral in another, a good match may be found because of the similarity in shape of the two sulci and their close proximity (as noted in Royackker’s study (Royackkers et al., 1999)). Presumably a mismatch would be found elsewhere in the brain— but the interpretation that the variability between the subjects was confined to this second area

810

FREE ET AL.

FIG. 6. Wireframe models of shape differences estimated by the principal component analysis of 12 landmarks in 116 cerebral hemispheres in 58 subjects (from Fig. 5). The wireframe models are shown in three orthogonal views. In the top row, the wireframes are of the landmark distributions at negative PC2 (mostly left hemispheres) and in the bottom row the wireframes are of the landmark distributions at positive PC2 (mostly right hemispheres). The right hemisphere landmarks were flipped about the midsagittal plane prior to the PC analysis to allow direct comparison between the two hemispheres.

would be erroneous. The sulcal matching of Lohmann is automatic and therefore reproducible and gives greater coverage of the brain but must be limited in its biological interpretation because of the lack of homology. The use of nonlinear normalization (as in Lohmann’s study above) allows more direct analysis and widespread comparison of subjects but must involve the alteration of relational structural differences and in ways which are not manifest. Thompson noted that some of their findings in comparing several sulci throughout the brain may have been attributed to the use of the Talaraich normalization process for bringing brains into a common space (Thompson et al., 1996). This process reduces variation close to the control points (anterior and posterior commissure) but has less effect on variation at distances from the control points (i.e., on the cortical surface). A similar point was noted by Royackkers who measured the residual variation in 6 sulci after nonlinear registration (Royackkers et al., 1999). The true relations between sulci may therefore be obscured in nontransparent ways dependent upon the registration process used. No correlations of shape differences with age were found in our study. Global atrophy which occurs during normal aging will be removed by the Procrustes fit and

in this study the limited number of landmarks may not be sufficient to detect global volume changes. In our data there was no correlation of increasing age with decreasing centroid size but given that our subjects were predominately young (mean age 33 years, maximum age 59 years) this is perhaps not surprising. Local age related atrophy, which may be postulated to alter the relative distances between landmarks, has been reported in cross-sectional studies [e.g., Raz et al., 1997) but longitudinal studies have found much less atrophy and speculate bias in cross-sectional studies due to some older subjects suffering atrophy because of preclinical dementias (Mueller et al., 1998). The visualization of local deformations which is possible with our approach may be of use in identifying local atrophic changes in older subjects or subtle early changes in degenerative conditions especially if more landmarks can be identified. We found no correlations of shape differences with gender. Size differences between males and females have been the only structural brain differences found with consistency in the literature apart from a recent study finding variations in corpus callosum cross-sectional area related to gender (Davatzikos and Resnick, 1998). However, Jancke et al., investigating landmarks about the corpus callosum speculate that gender dif-

MORPHOMETRICS OF THE ADULT BRAIN USING MRI

ferences are only due to size differences (Jancke et al., 1997). That is, increasing brain size is the driving force for hemispheric specialisation and as men tend to have bigger brains than women this is noted as a gender difference. Since the morphometric approach explicitly removes size dependency this removes the main difference between males and females. However, Ashburner also reported gender differences in an analysis of deformation fields in brain scan normalisation, finding more protruding occipital poles in males than females and greater protrusion of the frontal poles in females (Ashburner et al., 1998). This is at variance with our findings but may be explained by the differences in analysis techniques between the two studies. This type of deformation analysis, recently applied by Gaser et al. (1999), does consider the whole brain and can give an overall picture of the cortical variability. Gaser considered the differences between control subjects and patients with schizophrenia but did not provide any information about differences within the control group. This type of analysis still has theoretical difficulties which have not yet been resolved. For example, for statistical validity these techniques require a priori hypotheses—these are not always possible or desirable. Several of the studies discussed above have considered the variation in individual sulci but the form of analysis gives no sense of how the chosen sulci vary relationally to each other within any one brain. Our analysis is designed to consider relational differences and the wireframe visualisation makes these variations visually explicit. The whole brain grows as one connected structure, not each sulcus in isolation so the relation of sulci to each other must be important for understanding brain development. The brain coverage offered by widespread landmarks improves the completeness of interpretation, but with the loss of detail. The morphometric shape analysis only describes the variation in position of the chosen landmarks with the variation between landmarks only inferred as a linear extrapolation between those landmarks. If the number of landmarks does not adequately cover the structure of interest then only a partial description of the shape changes is possible. A further limitation in medical images is the identification of sufficient readily identifiable landmarks. The power of the analysis is dependent upon the extent to which they are homologous. As we have noted the correct labeling of substructures of the brain is the primary difficulty in this type of study in the brain. In the literature landmarks have been defined with a spectrum of confidence concerning the degree of homology (Marcus et al., 1996). There is potential for future analysis which weights the statistical analysis according to the confidence in the quality of the landmarks (Dryden and Mardia, 1998). Statistical developments are at present focusing on using lines, as opposed to points, and nonhomologous landmarks as

811

the basis for a statistical analysis (Bookstein, 1997; Dryden and Mardia, 1998). This is likely to make statistical shape analysis techniques more generally applicable. In addition these techniques are ideally formulated for studying developmental changes (as evidenced by the origin of statistical shape analysis within the evolutionary biology realm), which can be considered ontogenetically as well as phylogenetically. In summary the use of geometric morphometrics with landmarks on the surface of the cerebral cortex has some advantages over other types of analysis. The use of Procrustes fit for landmarks to remove location and size dependence is robust and has minimal impact on the subsequent interpretation of the data. The principle of attempting to match homologous points improves the biological significance of the results. An overall interpretation of the differences and their relation to each other can be found. The use of the principal component analysis gives some ranking to the types of shape differences seen. The major limitation is that there is limited coverage with landmarks. In the brain this could be improved with better anatomical labeling and more landmarks through the depth of the brain tissue rather than just at the surface. Even with limited numbers of landmarks differences in normal brains were identified. It is therefore possible that subtle differences in the brain due to pathological processes may be detectable. Abnormal brain development can alter the relative sulcal distribution in the cerebral hemisphres. While this may be obvious to visual inspection in some subjects, the ability to quantify sulcal distribution may assist in identifying more subtle forms of maldevelopment. ACKNOWLEDGMENTS The authors thank Dr. S. M. Sisodiya and Dr. F. G. Woermann for the use of their segmented MRI data. The financial support of the Medical Research Council (UK) and the National Society for Epilepsy is gratefully acknowledged.

REFERENCES Abell, F., Krams, M., Ashburner, J., Passingham, R., Friston, K., Frackowiak, R., Happe, F., Frith, C., and Frith, U. 1999. The neuroanatomy of autism: A voxel-based whole brain analysis of structural scans. Neuroreport 10(8): 1647–1651. Amunts, K., Schlaug, G., Schleicher, A., Steinmetz, H., Dabringhaus, A., Roland, P. E., and Zilles, K. 1996. Asymmetry in the human motor cortex and handedness. Neuroimage 4: 216 –222. Amunts, K., Jancke, L., Mohlberg, H., Steinmetz, H., and Zilles, K. 2000. Interhemispheric asymmetry of the human motor cortex related to handedness and gender. Neuropsychologia 38: 304 –312. Ashburner, J., Hutton, C., Frackowiak, R., Johnsrude, I., Price, C., and Friston, K. 1998. Identifying global anatomical differences: Deformation-based morphometry. Hum. Brain Mapp. 6: 348 –357. Blatter, D. D., Bigler, E. D., Gale, S. D., Johnson, S. C., Anderson, C. V., Burnett, B. M., Parker, N., Kurth, S., and Horn, S. D. 1995.

812

FREE ET AL.

Quantitative volumetric analysis of brain MR normative database spanning 5 decades of life. Am. J. Neuroradiol. 16: 241–251. Bookstein, F. L. 1989. Principal warps: Thin-plate splines and the decomposition of deformations. IEEE Trans. PAMI 11: 567–585. Bookstein, F. L. 1997. Shape and information in medical images: A decade of the morphometric synthesis. Comput. Vis. Image Understand. 66(2): 97–118. Buchanan, R. W., Vladar, K., Barta, P. E., and Pearlson, G. D. 1998. Structural evaluation of the prefrontal cortex in schizophrenia. Am. J. Psychiatr. 155: 1049 –1055. Buckley, P. F., Dean D., Bookstein, F. L., Friedman, L., Kwon, D., Lewin, J. S., Kamath, J., and Lys, C. 1999. Three-dimensional magnetic resonance-based morphometrics and ventricular dysmorphology in schizophrenia. Biol. Psychiatry 4: 562– 67. Chapman, L. J., and Chapman, J. P. 1987. The measurement of handedness. Brain Cogn. 6: 175–183. Davatzikos, C., and Resnick, S.M. 1998. Sex differences in anatomic measures of interhemispheric connectivity correlations with cognition in women but not men. Cerebral Cortex 8: 635– 640. DeQuardo, J. R., Keshavan, M. S., Bookstein, F. L., Bagwell, W. W., Green, W. D. K., Sweeney, J. A., Haas, G. L., Tandon, R., Schooler, N. R., and Pettegrew, J. W. 1999. Landmark-based morphometric analysis of first-episode schizophrenia. Biol. Psychiatry 45: 1321– 1328. Dryden, I. L., and Mardia, K. V. 1993. Multivariate shape analysis. Sankya 55(A): 460 – 480. Dryden, I. L., and Mardia, K. V. 1998. Statistical Shape Analysis Wiley and Sons, London. Duvernoy, H. 1991. The Human Brain. Springer Verlag, Vienna. Florack, L. M. J. 1993. The Syntactical Structure of Scalar Images. PhD thesis, University of Utrecht. Gaser, C., Volz, H. P., Kiebel, S., Riehemann, S., and Sauer, H. 1999. Detecting structural changes in whole brain based on nonlinear deformations—Application to schizophrenia research. Neuroimage 10(2): 107–113. Good, P. 1993. Permutation Tests: A Practical Guide to Resampling Methods for Testing Hypotheses. Springer-Verlag, New York. Goodall, C. R., and Mardia, K. V. 1991. A geometrical derivation of the shape density. Adv. Appl. Probability 23: 496 –514. Hutsler, J. J., Loftus, W. C., and Gazzaniga, M. S. 1998. Individual variation of cortical surface area asymmetries. Cerebral Cortex 8: 11–17. Ide, A., Rodriguez, E., Zaidel, E., and Aboitiz, F. 1996. Bifurcation patterns in the human Sylvian fissure: Hemispheric and sex differences. Cerebral Cortex 6(5): 717–725. Ide, A., Dolezal, C., Fernandez, M., Labbe, E., Mandujano, R., Montes, S., Segura, P., Verschae, G., Yarmuch, P., and Aboitiz, F. 1999. Hemispheric differences in variability of fissural patterns in parasylvian and cingulate regions of human brains. J. Comp. Neurol. 410: 235–242. Jack, C. R., Petersen, R. C., Xu, Y. C., Waring, S. C., O’Brien, P. C., Tangalos, E. G., Smith, G. E., Ivnik, R. J., and Kokmen, E. 1997. Medial temporal atrophy on MRI in normal aging and very mild Alzheimer’s disease. Neurology 49: 786 –794. Jancke, L., Schlaug, G., Huang, Y., and Steinmetz, H. 1994. Asymmetry of the planum parietale. NeuroReport 5: 1161–1163. Jancke, L., Staiger, J. F., Schlaug, G., Huang, Y., and Steinmetz, H. 1997. The relationship between corpus callosum size and forebrain volume. Cerebral Cortex 7: 48 –56. Kendall, D. G. 1984. Shape manifolds, procrustean metrics and complex projective spaces. Bull. Lond. Math. Soc. 16: 81–121. Kennedy, D. N., Lange, N., Makris, N., Bates, J., Meyer, J., and Caviness, V. S. 1998. Gyri of the human neocortexan MRI-based analysis of volume and variance. Cerebral Cortex 8: 372–384.

Kido, D. K., Lemay, M., Levinson, A. W., and Benson, W. E. 1980. Computed tomographic localisation of the precentral gyrus. Radiology 135: 373–377. Kulynych, J. J., Vladar, K, Fantie, B. D., Jones, D. W., and Weinberger, D. R. 1995. Normal asymmetry of the planum temporale in patients with schizophrenia. Br. J. Psychiatry 166: 742–749. Kulynych, J. J., Vladar, K., Jones, D. W., and Weinberger, D. R. 1996. Superior temporal gyrus volume in schizophreniaa study using MRI morphometry assisted by surface rendering. Am. J. Psychiatry 15: 350 –56. Le Goualher, G., Argenti, A., Duyme, M., Baare, W. F. C., Hulshoff Pol, H. E., Boomsma, D. I., Zouaoui, A., Barillot, C., and Evans, A. C. 2000. Statistical sulcal shape comparisons: Application to the detection of genetic encoding of the central sulcus shape. Neuroimage 11: 564 –574. Lohmann, G., von Cramon, D. Y., and Steinmetz, H. 1999. Sulcal variability of twins. Cerebral Cortex 9: 754 –763. Manceaux-Demiau, A., Bryan, R. N., and Davatzikos, C. 1998. A probabilistic ribbon model for shape analysis of the cerebral sulci: Application to the central sulcus. J. Comput. Assist. Tomogr. 22(6): 962–971. Marcus L. F., Corti M., Loy A., Naylor G. J. P., and Slice D. (Eds.) 1996. Advances in Morphometrics, Nato ASI Series. Plenum Press, New York. Maudgil, D. D., Free, S. L., Sisodiya, S. M., Lemieux, L., Woermann, F. G., Fish, D. R., and Shorvon, S. D. 1998. Identifying homologous anatomical landmarks on reconstructed magnetic resonance images of the human cerebral cortical surface. J. Anat. 193: 559 –571. Mueller, E. A., Moore, M. M., Kerr, D. C. R., Sexton, G., Camicioli, R. M., Howieson, D. B., Quinn, J. F., and Kaye, J. A. 1998. Brain volume preserved in healthy elderly through the eleventh decade. Neurology 51: 1555–1562. O’Higgins, P., and Jones, N. 1998. Facial growth in Cercocebus torquatusan application of three-dimensional geometric morphometric techniques to the study of morphological variation. J. Anat. 193: 251–272. O’Higgins, P. 2000. The study of morphological variation in the hominid fossil record: Biology, landmarks and geometry. J. Anat. 197: 103–120. Penhune, V. B., Zatorre, R. J., MacDonald, J. D., and Evans, A. C. 1996. Interhemispheric anatomic differences in human primary auditory cortex: Probabilisitic mapping and volume measurement from magnetic resonance scans. Cerebral Cortex 6(5): 661– 672. Raz, N., Gunning, F. M., Head, D., Dupuis, J. H., McQuain, J., Briggs, S. D., Loken, W. J., Thornton, A. E., and Acker, J. D. 1997. Selective aging of the human cerebral cortex observed in vivo: Differential vulnerability of the prefrontal gray matter. Cerebral Cortex 7: 268 –282. Rohlf, F. J. 1999. Shape statistics: Procrustes superimpositions and tangent space. J. Class. 16: 197–223. Royackkers, N., Desvignes, M., Fawal, H., and Revenu, M. 1999. Detection and statistical analysis of human cortical sulci. Neuroimage 10: 625– 641. Rumsey, J. M., Donohue, B. C., Brady, D. R., Nace, K., Giedd, J. N., and Andreason, P. 1997. A magnetic resonance imaging study of planum temporale asymmetry in men with developmental dyslexia. Arch. Neurol. 54: 1481–1489. Sandor, S., and Leahy, R. 1997. Surface-based labelling of cortical anatomy using a deformable atlas. IEEE Trans. Med. Imag. 16(1): 41–54. Sisodiya, S. M., Free, S. L., Stevens, J. M., Fish, D. R., and Shorvon, S. D. 1995. Widespread cerebral structural changes in patients with cortical dysgenesis and epilepsy. Brain 118: 1039 –1050.

MORPHOMETRICS OF THE ADULT BRAIN USING MRI Sisodiya, S. M. 1996. Qualitative and quantitative analysis of MRI data from patients with epilepsy. PhD thesis. University of London, UK. Sorlie, C., Bertrand, O., Yvert, B., Froment, J.-C., and Pernier, J. 1997. Matching of digitised brain atlas to magnetic resonance images. Med. Biol. Eng. Comput. 35: 239 –245. Thompson, P. M.., Schwartz, C., Lin, R. T., Khan, A. A., and Toga, A. W. 1996. Three-dimensional statistical analysis of sulcal variability in the human brain. J. Neurosci. 16(13): 4261– 4274. Thompson, P. M.., Moussai, J., Zohoori, S., Goldkorn, A., Khan, A. A., Mega, M. S., Small, G. W., Cummings, J. L., and Toga, A. W. 1998. Cortical variability and asymmetry in normal aging and Alzheimer’s disease. Cerebral Cortex 8: 492–509. Tramo, M. J., Loftus, W. C., Thomas, C. E., Green, R. L., Mott, L. A., and Gazzaniga, M. S. 1995. Surface area of human cerebral cortex and its gross morphological subdivisionsin vivo measurements in monozygotic twins suggest differential hemisphere effects of genetic factors. J. Cogn. Neurosci. 7(2): 292–301. Westbury, C. F., Zatorre, R. J., and Evans, A. C. 1999. Quantifying variability in the planum temporale: A probability map. Cerebral Cortex 9: 392– 405.

813

White, L. E., Andrews, T. J., Hulette, C., Richards, A., Groelle, M., Paydarfar, J., and Purves, D. 1997a. Structure of the human sensorimotor system I: Morphology and cytoarchitecture of the central sulcus. Cerebral Cortex 7: 18 –30. White, L. E., Andrews, T. J., Hulette, C., Richards, A., Groelle, M., Paydarfar, J., and Purves, D. 1997b. Structure of the human sensorimotor system II: Lateral symmetry. Cerebral Cortex 7: 31– 47. Witelson, S. F., and Kigar, D. L. 1992. Sylvian fissure morphology and asymmetry in men and women: Bilateral differences in relation to handedness in men. J. Comp. Neurol. 323: 326 –340. Woermann, F. G., Free, S. L., Koepp, M. J., Ashburner, J., and Duncan, J. S. 1999. Voxel-by-voxel comparison of automatically segmented cerebral gray matter—A rater-independent comparison of structural MRI in patients with epilepsy. Neuroimage 10: 373– 384. Zilles, K., Dabringhaus, A., Geyer, S., Amunts, K., Qu, M., Schleicher, A., Gilissen, E., Schlaug, G., and Steinmetz, H. 1996. Structural asymmetries in the human forebrain and the forebrain of non-human primates and rats. Neurosci. Biobehav. Rev. 20(4): 593– 605.