A Method for Assessing the Accuracy of Intersubject Registration of the Human Brain Using Anatomic Landmarks

NeuroImage 9, 250–268 (1999) Article ID nimg.1998.0397, available online at http://www.idealibrary.com on

A Method for Assessing the Accuracy of Intersubject Registration of the Human Brain Using Anatomic Landmarks Igor D. Grachev,*,†,§,¶ Dmitriy Berdichevsky,†,¶ Scott L. Rauch,*,†,§,¶ Stephan Heckers,*,§ David N. Kennedy,‡,\ Verne S. Caviness,‡,\ Nathaniel M. Alpert†,¶ *Department of Psychiatry, †Department of Radiology, and ‡Department of Neurology, Harvard Medical School, §Psychiatric Neuroimaging Research Group, Department of Psychiatry, ¶PET Imaging Laboratory, Division of Nuclear Medicine, Department of Radiology, and the \Center for Morphometric Analysis, Department of Neurology, Massachusetts General Hospital, Boston, Massachusetts 02114 Received June 2, 1998

Several groups have developed methods for registering an individual’s 3D MRI by deforming a standard template. This achievement leads to many possibilities for segmentation and morphology that will impact nuclear medical research in areas such as activation and receptor studies. Accordingly, there is a need for methods that can assess the accuracy of intersubject registration. We have developed a method based on a set of 128 anatomic landmarks per hemisphere, both cortical and subcortical, that allows assessment of both global and local transformation accuracy. We applied our method to compare the accuracy of two standard methods of intersubject registration, AIR 3.0 with fifth-order polynomial warping and the Talairach stereotaxic transformation (Talairach and Tournoux, 1988). SPGR MRI’s (256 3 256 3 160) of six normal subjects (age 18–24 years) were derformed to match a standard template volume. To assess registration accuracy the landmarks were located on both the template volume and the transformed volumes by an experienced neuroanatomist. The resulting list of coordinates was analyzed graphically and by ANOVA to compare the accuracy of the two methods and the results of the manual analysis. ANOVA performed over all 128 landmarks showed that the Woods method was more accurate than Talairach (left hemisphere F 5 2.8, P F 0.001 and right hemisphere F 52.4, P F 0.006). The Woods method provided a better brain surface transformation than did Talairach (F 5 18.0, P F 0.0001), but as expected there was a smaller difference for subcortical structures and both had an accuracy F1 mm for the majority of subcortical landmarks. Overall, both the Woods and Talairach method located about 70% of landmarks with an error of 3 mm or less. More striking differences were noted for landmark accuracy I1 mm, where the Woods method located about 40% and Talairach about 23%. These results demonstrate that this anatomically based assessment method can help evaluate new methods of intersubject registration and should be a helpful tool in appreciating regional differ1053-8119/99 $30.00 Copyright r 1999 by Academic Press All rights of reproduction in any form reserved.

ences in accuracy. Consistent with expectation, we confirmed that the Woods nonlinear registration method was more accurate than Talairach. Landmarkbased anatomic analyses of intersubject registration accuracy offer opportunities to explore the relationship among structure, function and architectonic boundaries in the human brain. r 1999 Academic Press

INTRODUCTION Studies of cognitive function or abnormal structure must account for the normal variation of brain anatomy. Until recently, all morphometric studies relied on neuroanatomic expertise to overcome the striking variations that exist across individuals with respect to whole brain anatomy and particularly to the sulcal anatomy of neocortex (Rademacher et al., 1993; Rajkowska and Goldman-Rakic, 1995; Paus et al., 1996a, 1996b; Thompson et al., 1996). Accordingly, hallmarks of this type of research are processes that require time consuming manual segmentation and anatomic localization, methods that become more burdensome as the slice thickness of modern imaging techniques decreases. Meanwhile, functional imaging studies that evolved over the past decade have approached the problem of anatomic variability from another perspective, one in which a brain is transformed by a computer algorithm to match a particular template or atlas. The best known example of this approach is the Talairach transformation (Talairach and Tournoux, 1988), that matches brain size and orientation, but does not attempt to produce a detailed match of the neocortical sulcal patterns. Recently, investigators have begun to apply more complex, nonlinear, intersubject registration methods in the hope of achieving more accurate matching of individual brains to a standard anatomic template. In principle, the method of intersubject registration offers the possibility of eliminating the need for manual procedures. The basic idea of this approach is to label/segment/measure

250

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

a standard brain image volume (i.e., the template) once and then to automatically transfer the information to the corresponding locations in another individual’s brain images via intersubject registration. Several groups have reported methods for intersubject registration using warping algorithms (Bookstein, 1989; Evans et al., 1991; Friston et al., 1991; Miller et al., 1993; Grenander and Miller, 1994; Toga, 1994; Collins et al., 1995; Davatzikos, 1996; Schormann et al., 1996; Thompson and Toga, 1996; Warfield et al., 1996; Iosifescu et al., 1997; Thompson et al., 1997; Woods et al., 1998a, 1998b), but standard methods for evaluating the global and regional accuracy of the transformation have not been fully developed. This paper describes a method for assessing the accuracy of intersubject brain registration and exemplifies the method by comparing the accuracy of two intersubject registration schemes. The approach we have taken is based on the measurement of anatomic landmarks that are easily identified in high resolution MRI images of the human brain. One of the novel aspects of this work is the method of identifying neocortical landmarks that depends on the intersection of standardized lines and planes with the sulcal anatomy. Landmark-based measurement has a long history in neuroanatomic research and other areas, as well. Analysis of landmark location has been used in applications such as forensics (Dongsheng and Yuwen, 1993; Georg, 1993), computer-assisted neurosurgery (Kikinis et al., 1996), anthropological study (Novotny et al., 1993), MRI-based morphological analyses of whole brain (Bookstein, 1994; Filipek et al., 1994), landmarkbased registration and measurement of MRI (Arndt et al., 1996), MRI-based morphometric topographic parcellation of specific human neocortical subunits in normal control subjects (Rademacher et al., 1992; Caviness et al., 1996; Grachev et al., 1996; Wible et al., 1997), and in the assessment of psychiatric disorders (Grachev et al., 1998). Landmarks provide size and shape information. They can reflect the uniqueness of brain structure, especially topographic specificity of the sulcal and gyral pattern of the neocortex. While techniques that use landmark information are already used in other fields, their application to quality assessment of intersubject registration is still relatively new, especially with respect to nonlinear plastic transformation technologies. We used anatomic landmark locations to address several questions: (1) Can these anatomical landmarks be used to assess the global quality of two different approaches in MRI intersubject registration: (a) The nonlinear transformation method of Woods AIR 3.0 (Woods et al., 1998a, 1998b) and (b) The piece-wise linear Talairach stereotaxic transformation method (Talairach and Tournoux, 1988). (2) Is the anatomical landmark method sufficiently sensitive to evaluate the

251

regional differences between these two registration methods? METHODS MRI Data MRI data were acquired using an SPGR sequence (flip angle 30°, TE minimum, TR 31, scan time 10 min). The image volume was reconstructed as a series of 160 coronal planes, with 256 3 256 pixels each. Voxel size was 0.937 mm in plane, with a slice thickness of 1.5 mm. Subjects Seven normal subjects (age range 18–24 years) were originally recruited for another study and their MRI scans were later used in this evaluation. Subjects were college students who reported that they were in good health, not addicted to drugs or alcohol, and without any history of neurological disease, major medical illnesses, or psychiatric disorder. Intersubject Registration Intersubject registration was performed by two methods, a Talairach stereotaxic transformation and a nonlinear registration method developed by Woods et al. Talairach stereotaxic transformation was performed manually by selection of key anatomic landmarks (anterior commissure (AC), posterior commissure (PC), and in each hemisphere frontal pole, occipital pole, base of temporal lobe, vertex, and lateral pole), followed by piece-wise linear transformation according to Talairach and Tournoux (1988). After Talairach transformation, the brain was oriented such that the origin of coordinates was centered at the AC in the midsagittal plane. Nonlinear registration was performed using AIR 3.0 software supplied by Woods et al. (1998a, 1998b). The intersubject registration method of Woods and colleagues is an extension of their previous intrasubject registration algorithm, with the major modification being the spatial transformation model which has been extended to allow polynomial coordinate transformation functions. Our application employed a fifth-order polynomial transformation. A typical normal scan was designated as the reference scan and manually transformed as described above so that it was standardized in Talairach space. Each scan was manually edited to remove the extracranial structures (scalp, skull, and meninges), the brain stem (except the ventral diencephalon to make AC and PC points visible), and cerebellum. Subsequent intersubject registration was performed to deform the subject data so that it was in registration with the reference scan.

252

GRACHEV ET AL.

Analysis of Registration Accuracy Registration error was defined to be the Euclidean distance between the location of a reference scan landmark and the homologous point after nonlinear registration. Landmarks were uniquely defined by a procedure detailed below. The location of each landmark was determined by an experienced neuroanatomist and tabulated as a (x,y,z) triplet for each hemisphere, yielding a table for each subject and each method with 128 landmarks and 6 coordinates. The data table for each subject was combined with the corresponding data table from the reference scan to compute the mean error for each landmark. Differences in registration error were analyzed with analysis of variance (ANOVA), using the general linear model (JMP, SAS Institute, Cary, NC). The outcome variable was taken to be the absolute distance between corresponding landmarks. Left and right hemispheres were analyzed separately. In these ANOVAs, registration method, landmark number, region, and subject were used as explanatory variables. Additional details are given below. Anatomic Landmark Analyses To aid the evaluation of the accuracy of intersubject registration we developed a set of well-defined anatomic landmarks to serve as a metric for the comparison. Specific criteria were used for the landmark selection: (1) Wide distribution of landmarks throughout the brain; (2) Characterization of the topographic uniqueness of a whole individual brain including neocortex; (3) Relatively simple and accurate identification of all landmarks by any investigator for all MRI images in a manageable period of time. The landmarks are defined relative to a coordinate system in a 3-D space (the coronal plane is XZ, the sagittal plane is YZ, the transaxial plane is XY), whose origin is at the AC, with axes determined by the intersection of the two horizontal lines [CC (corpus callosum), AC–PC], one vertical AC line, and the midsagittal plane. The overall organization of the system is shown in Fig. 1. Definition of Anatomic Reference Planes Midsagittal planes [represents most medial sagittal planes for both hemispheres] (Figs. 1A–1E) were defined cutting the AC on the right (right midsagittal plane) and left (left midsagittal plane). These planes provide information about brain shape (intersection of hemispheric margins with two horizontal lines (CC, AC–PC) and two vertical lines (AC, PC), the shape of corpus callosum (anterior and posterior points, posterior angle and tip of genu, fornix junction, inferior notch and posterior tip of splenium, splenium center), and localization of some neocortical regions (cingulate

and paracingulate gyri; parietooccipital, calcarine, subparietal, and cuneal fissures). Five axial (i.e., horizontal) planes are located by intersection with a line drawn perpendicular (i.e., vertical) to the AC–PC line, through the AC point, as illustrated in Figs. 1A–1E. For example, the axial AC–HM (AC-hemispheric margin) plane (Fig. 1A) was defined as the intersection of the AC vertical line and the hemispheric margin, using either a midsagittal or coronal plane. The axial AC–HM plane was used for characterization of topographic specificity in localization of precentral, central, and postcentral fissures on the top level of each individual brain. Axial AC–PaCiG (AC–paracingulate gyrus) plane (Fig. 1B) was identified as the intersection of the AC vertical line on midsagittal plane and the point along this line picked at the visual center of paracingulate gyrus. The axial AC–PaCiG plane was chosen for localization of neocortical fissures (superior frontal, precentral, central, postcentral, cingulate, intraparietal, and Jensen’s fissures). Axial CC plane (Fig. 1C) was defined as the horizontal CC line through the posterior angle of genu of corpus callosi (point No. 80) on right and left midsagittal planes. We used the axial CC plane for the landmark-based description of subcortical structures (caudate nucleus, putamen, lateral ventricle) and some regions of neocortex (insula, Sylvian fussure). Axial AC–PC plane (Fig. 1D) was determined by the AC–PC line on any midsagittal plane. The axial AC–PC plane was used for characterization of the same anatomical structures as described on the Axial CC plane. Axial Chiasmatic plane (Fig. 1E) was identified as the chiasma opticum point (point No. 14) on the Coronal AC plane. This is a very important plane for characterization of the frontal orbital cortex (medial, lateral and arcuate orbital sulci, and olfactory fissure). Only one coronal AC plane (Fig. 1D) was defined as the AC point (point No. 4), using either a coronal or midsagittal plane. The coronal AC plane provides valuable information about some subcortical structures (caudate nucleus, putamen, globus pallidus, chiasma opticum, infundibulum, ventricles) and insula. Lateral-sagittal planes (Fig. 1D) were determined as the point on Axial AC–PC plane picked at the halfdistance between right insula posterior tip and hemisphere margin (right lateral sagittal plane) and the same distance on the left (left lateral sagittal plane). These planes are very important in topographic localization of inferior frontal, precentral, central, postcentral, parietal, and temporal subregions of neocortex. The Anatomic Landmark Protocol for the Reference Brain (See Table, Appendix) The anatomic landmark protocol for the reference brain included 256 landmarks for both hemispheres (80 landmarks for subcortical structures and 176 landmarks for neocortex). Ninety-two percent of all land-

FIG. 1. (A–E) The overall organization of the system of anatomic reference planes. Midsagittal planes (represents most medial sagittal planes for both hemispheres) were defined cutting the AC on the right (right midsagittal plane) and left (left midsagittal plane). Five axial (i.e., horizontal) planes are located by intersection with a line drawn perpendicular (i.e., vertical) to the AC-PC line, through the AC point. Axial AC-HM (AC-hemispheric margin) plane (Fig. 1A) was defined as the intersection of the AC vertical line and the hemispheric margin. Axial AC-PaCiG (AC-paracingulate gyrus) plane (Fig. 1B) was identified as the intersection of the AC vertical line on midsagittal plane and the point along this line picked at the visual center of paracingulate gyrus. Axial CC plane (Fig. 1C) was defined as the horizontal CC line through the posterior angle of genu of corpus callosi [point No. 80] on right and left midsagittal planes. Axial AC-PC plane (Fig. 1D) was determined by the AC-PC line on any midsagittal plane. Axial Chiasmatic plane (Fig. 1E) was identified as the chiasma opticum point [point No. 14] on the Coronal AC plane. Coronal AC plane (Fig. 1D) was defined as the AC point [point No. 4], using either a coronal or midsagittal plane. Lateral Sagittal planes (Fig. 1D) were determined as the point on Axial AC-PC plane picked at the half-distance between right insula posterior tip and hemisphere margin (right lateral sagittal plane) and the same distance on the left (left lateral sagittal plane). 253

254

GRACHEV ET AL.

marks were distributed within 10 planes: Coronal AC plane (28 landmarks), Axial AC–HM plane (12 landmarks), Axial AC–PaCiG plane (24 landmarks), Axial CC plane (24 landmarks), Axial AC–PC plane (22 landmarks), Axial chiasmatic plane (16 landmarks), two midsagittal planes (72 landmarks), two lateral sagittal planes (38 landmark). Only 20 coronal landmarks were distributed along the Y axis and not located in any specific plane.

Error level

Reliability of Landmarks

1–2 mm

Intraobserver differences in anatomic landmarks for the reference brain were assessed by repeated measurement of the reference brain after an interval of 1 month. The root mean squared (RMS) error between landmarks was 1.6 mm for both left and right hemispheres. Preliminary analysis demonstrated heteroscedasticity for cortical versus subcortical landmarks.

2–3 mm

TABLE 1 Magnitude of Error for Anatomic Landmarks, by Region (Woods Method)

,1 mm

3–4 mm

RESULTS The Difference between the Two Transformation Algorithms We performed ANOVAs using the general linear model (JMP, SAS Institute) to test whether our method could detect differences in accuracy between the Woods algorithm and the Talairach transformation. The results were: 1. The two transformation algorithms (Talairach, Woods) differ significantly in the localization of 128 landmarks (ANOVA with algorithm—Woods or Talairach—as the main effect, with subject treated as a repeated measure; F 5 2.8, df 5 11, P , 0.001 for left hemisphere and F 5 2.4, df 5 11, P , 0.006 for right hemisphere). 2. The two transformation algorithms differ significantly across the 128 landmarks (ANOVA with two main effects: landmark, analysis (repeated for each subject); F 5 13.6, df 5 138, P , 0.0001 for left hemisphere and F 5 11.4, df 5 138, P , 0.0001 for right hemisphere). 3. The two transformation algorithms differ significantly between the two groups of regions (cortical, 88 landmarks; subcortical, 40 landmarks) (ANOVA with two main effects: region, analysis (repeated for each subject); F 5 18.0, df 5 12, P , 0.0001 for left hemisphere and F 5 15.1, df 5 12, P , 0.0001 for right hemisphere). 4. The two transformation algorithms differ significantly across the 128 landmarks when they are grouped into cortical and subcortical regions (ANOVA with three main effects: region, analysis (repeated for each subject), landmark (within measure for region); F 5 13.6, df 5 138, P , 0.0001 for left hemisphere

4–5 mm

5–10 mm

10–15 mm

Anatomic region (n) Cerebrum anterior (3); Cerebrum superior (4); Lateral ventricle (4); Putamen (12); Caudate (10); Globus pallidus (2); Thalamus (2); Corpus callosum (12); Anterior commissure (2); Posterior commissure (2); Chiasma opticum (2); Infundibulum (2); Cingulate cortex (2); Heschl’s Gyrus (2); Insula Cortex (2) Lateral ventricle (7); Lenticulate (10); Corpus callosum (10); Putamen (2); Fornix (2); Lateral geniculate nucleus (2); Cingulate cortex (10); Heschl’s gyrus (2); Insula cortex (6); Inferior frontal cortex (2); Orbital frontal cortex (2); Cerebrum anterior (1) Cerebrum posterior (4); Lateral ventricle (2); Lenticulate (2); Putamen (2); Corpus callosum (2); Cingulate cortex (2); Insula cortex (4); Sylvian fissure (2); Precentral fissure (2); Central fissure (6); Postcentral fissure (4); Inferior frontal cortex (4); Orbital frontal cortex (6); Occipital cortex (4); Parietal cortex (6); Temporal cortex (4) Sylvian fissure (4); Precentral fissure (2); Central fissure (4); Postcentral fissure (2); Inferior frontal cortex (2); Orbital frontal cortex (4); Occipital cortex (2); Parietal cortex (2) Precentral fissure (6); Postcentral fissure (6); Orbital frontal cortex (2); Occipital cortex (8); Parietal cortex (4); Temporal cortex (2) Lateral ventricle [posterior] (2); Precentral fissure (2); Central fissure (2); Postcentral fissure (2); Inferior frontal cortex (4); Superior frontal fissure (4); Orbital frontal cortex (2); Occipital cortex (4); Parietal cortex [Jensen’s fissure] (2); Temporal cortex (2) Lateral ventricle [posterior] (2); Parietal cortex [Jensen’s fissure] (2)

Note. (n) The number of landmarks on the different planes for each anatomic region. Overall, there was a smaller error for subcortical landmarks vs cortical landmarks. High level of error was associated with significant individual variability of certain anatomic regions (e.g., occipitoparietal region, inferior frontal gyrus, orbital frontal cortex, and precentral, central and postcentral sulci).

and F 5 11.4, df 5 138, P , 0.0001 for right hemisphere). Figure 2 shows the cumulative fraction of landmark as a function of landmark error for the two methods. Both the Woods and Talairach method located about 70% of landmarks with an error of 3 mm or less. The curves were most different for error ,2 mm. More striking differences were noted for landmark accuracy #1 mm, where the Woods method located about 40% and Talairach about 23%. DISCUSSION The method for assessing intersubject registration of brain volumes described above has several virtues, including that its result is easily understood in the context of traditional manual segmentation methods.

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

255

FIG. 2. The cumulative fraction of landmarks as a function of landmark error for the two methods.

Human expertise was used as a gold standard. Intrarater performance on repeated measurement showed good reliability, with an RMS error of 1.6 mm for both the left and right hemisphere. Error in automatic registration was defined by localization of the homologous landmarks after nonlinear registration and was determined by the manual assessment procedure for different subcortical and cortical regions (gyri, sulci). Despite the development of various transformation methodologies, there have been only limited attempts so far to evaluate the accuracy of these methods. Previous reports of analysis by landmark location were validated using small numbers of landmarks, distributed mostly throughout the corpus callosum, basal ganglia, or brain stem, and applied primarily for linear nonplastic transformation algorithms (Neelin et al., 1993; Arndt et al., 1996). In one study, 27 subcortical landmarks were chosen from the midsagittal plane and were used for the 26 rotations based on the average angular deviation as part of the registration procedure (Arndt et al., 1996). None of these subcortical landmarks exceeded 3 mm in error, while most (16/27) were ,1 mm of mean error. Three-dimensional simulated PET images generated from MRI were used by another research group (Neelin et al., 1993) to validate a multimodality registration technique based on the identification of internal anatomical landmarks. Their registration errors ranged from 1.0 mm at the brain center to 2.8 mm in each dimension at the brain surface. Validation of their nonlinear elastic transformation

(Woods et al., 1998a, 1998b) accomplished the landmarkbased assessment by mathematically projecting each traced landmark point from the three-dimensional renderings back into the original three-dimensional volume of data, and then onwards into the final common space, using the particular set of registration parameters. The landmarks chosen for that assessment were the precentral and postcentral gyri, parietooccipital and calcarine sulci, superior temporal gyri, AC and PC points. Our approach shares some similarities with that of Woods et al. (1998a, 1998b); both reflect an anatomical orientation relying upon topographical characterization of the individual brain volume, especially neocortex. The current method represents an extension of the work by Woods et al. in that we sought to address a broader array of gyral and sulcal landmarks, which more fully characterize the brain surface. An anatomically based approach generally works very well if selected landmarks are identified by an experienced neuroanatomist. Inaccuracies in landmark localization may introduce additional errors, regardless of the anatomical scheme employed. In the present study, we developed an anatomically based assessment method to compare the accuracy of two standard methods of intersubject registration (AIR 3.0 and the Talairach stereotaxic transformation). We used a set of 128 anatomic landmarks per hemisphere widely distributed throughout cortical and subcortical regions. This method allowed us to assess both global and local

256

GRACHEV ET AL.

landmark accuracy. As expected we found that the Woods method provided a better brain surface transformation than did Talairach, and that there was a smaller difference between Woods and Talairach for subcortical structures; both methods had an accuracy ,1 mm for the majority of subcortical landmarks. As were shown, both methods might be used for characterization of subcortical structures (basal ganglia, thalamus, ventricles system, and corpus callosum) with a high level of accuracy. Overall, both the Woods and Talairach method located about 70% of landmarks with an error of 3 mm or less. The curves were most different for an error ,2 mm. More striking differences were noted for landmark accuracy #1 mm, where Woods located ,40% and Talairach ,23%, especially for landmarks presented throughout the neocortex. As we observed, the Woods method might be superior for analyses of morphofunctional differences within small neocortical areas, where the Talairach method is relatively insensitive and can introduce additional error. The Woods method was found to reduce, but not eliminate the residual anatomic variability of brain surface regions with significant individual and interhemispheric variations (e.g., occipitoparietal region, inferior frontal gyrus, orbital frontal cortex, and precentral, central, and postcentral sulci). We noted above that we observed different error variability of localization for cortical and subcortical landmarks, a finding that contradicts the assumption of homoscedasticity for analysis of variance calculations. The ANOVAs treating all 128 landmarks as a single group are most susceptible; while the ANOVAs stratifying landmarks into cortical and subcortical groups are less susceptible to this problem. Nevertheless, one would expect that heteroscedasticity would be more likely to blunt the ability of ANOVA to reveal differences between methods, therefore the P values associated with the analyses might be conservative. In this study, intersubject registration was achieved by matching of each individual brain to a standard anatomic template. By this approach, a standard brain image volume (i.e., the template) was described once and then automatically transferred to the corresponding locations in another individual’s brain images. We used a single normal MRI brain volume as the reference template for intersubject registration. As was noted by Woods et al., this procedure might not be ideal. It is important to consider that future studies should employ an average template (i.e., atlas) rather than a specific MRI volume. Theoretically, this average atlasbased approach in intersubject registration should provide a superior anatomical correspondence for each individual brain. It may also prove plausible to use this type of method to enact automated morphometric analysis. Specifically, once landmarks are defined on an intersubject

averaged template or a probabilistic atlas-based template (Thompson et al., 1997), they can be translated to the native space of each individual subject via backtransformation. This would greatly enhance the efficiency of morphometric studies involving neuropsychiatric disorders. However, the validity of such an approach remains to be tested. Nonlinear intersubject registration algorithms such as Wood’s AIR 3.0 may offer distinct advantages over linear methods for determining average morphology in groups of subjects. Nevertheless, problems remain with detailed registration of complex gyral and sulcal patterns. Problems such as mismatch due to extra branching, doubling, or deletion of gyri cannot be addressed in this approach. In fact, such instances increase the cost function and may cause small errors in the transformation. The present data are limited for several reasons, including small sample size and certain intrinsic mismatches of brain to template (e.g., missing section of gyrus or different topographic sulcal pattern). Although striking variations exist across individuals with respect to whole brain anatomy, and specifically with regard to sulcal anatomy of neocortex, the general approach of stereotactic Talairach transformation completely ignores this, the natural pattern of topographic interindividual variation of brain surface anatomy, and may be attended by substantial morphological and functional mismatching. Fully automated, hierarchical, intersubject nonlinear registration methods using elastic warping transformation models (Woods et al., 1998a, 1998b) provide better matching for distinct cortical features (gyri, sulci) (Collins and Evans, 1997). In general, intersubject registration might minimize the issue of dissimilarity in brain surface between different subjects that conserves the principle of topographic uniqueness of the individual brain. One might also criticize the basis of this work: Intersubject registration methods, such as that of Woods et al. should not be expected to align the landmarks used in this paper because they were not designed to do so. Furthermore, other methods are available, or could be devised, that could exactly match a set of landmark points across a group of subjects. This view is correct in a technical sense; but it is also rather narrow and impractical. The idea that one can design a registration algorithm that will align a specific set of points or curves is appealing. However, doing so would in no way guarantee that other points in the brain would be mapped in a sensible way. In fact, this is a well-known pitfall when constructing high order interpolation algorithms for a continuous mathematical function,—i.e., The interpolated function tends to oscillate in the region between the control points or knots. The approach of Woods et al. and others avoids this problem by using a low-order polynomial mapping to bring a given

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

subject’s brain into registration with a template. The intention is to reduce the residual anatomic variability of homologous voxels in the images across a cohort of subjects; it will be a very poor and disappointing registration method that will not reduce the distance of homologous points in the template and individual registrations. Our results demonstrate that this anatomically based assessment method can help evaluate new methods of intersubject registration and should be a helpful tool in appreciating regional differences in accuracy. Consistent with expectation, we confirmed that the Woods nonlinear registration method was more accurate than Talairach. Landmark-based anatomic analyses of intersubject registration accuracy offer opportunities to explore the relation of structure, function, architectonic boundaries in the human brain histologic labeling, and can be used in future research as a guide for the generation of anatomical atlases, automated segmentation and cortical parcellation schemes, analyses of structural variability in the human brain, and modeling neuropsychiatric disorders. ACKNOWLEDGMENTS Dr. Grachev was supported in part by a NIH National Research Service Award in Radiological Sciences (5T32CA09362-16) from the National Cancer Institute, Bethesda, MD. Dr. Rauch was supported in part by Grant MH01215 from NIMH, and a NARSAD Young Investigator Award.

APPENDIX Anatomic Landmark Protocol Table 2 lists the coordinates of the landmarks used in this study. The protocol for locating the landmarks is given below. Definitions of anatomic landmarks (these landmark numbers correspond to the same numbers as the landmark protocol): (A) Coronal landmarks. (localization of landmarks Nos. 4–17 presented on Fig. 3): 1. Lateral ventricle anterior was defined as the most anterior point which was picked at the visual center of the anterior horn of lateral ventricles (the first appearance of the anterior horn of lateral ventricles on the coronal view); 2. Temporal lobe anterior was defined as the most anterior point picked at the visual center of temporal lobe (the first appearance of temporal lobe on the coronal view); 3. Isthmus of temporal lobe was defined as the most anterior point on the coronal view where temporal lobe connects with frontal lobe (must correspond to the most posterior coronal plane of temporal lobe); 4. Anterior Commissure corresponds to the midpoint of decussation of the anterior commissure on the coronal AC plane (the point of origin of the coordinate system);

257

5. Lateral ventricle superior was defined as the most superior point of lateral ventricles on the coronal AC plane; 6. Lateral ventricle inferior was defined as the most inferior point of lateral ventricles on the coronal AC plane; 7. Caudate nucleus superior was defined as the most superior point of the head of caudate on the coronal AC plane; 8. Caudate nucleus inferior was defined as the most inferior point of the head of caudate on the coronal AC plane; 9. Putamen superior was defined as the most superior point of putamen on the coronal AC plane; 10. Putamen inferior was defined as the most inferior point of putamen on the coronal AC plane; 11. Lenticulate superior notch was defined as the most superior point of lamina medullaris lateralis on the coronal AC plane (lamina medullaris lateralis subdivides lenticulate into two parts—globus pallidus and putamen). Select intersection of superior tip of lamina medullaris lateralis with the capsula interna; 12. Lenticulate inferior notch was defined as the most inferior point of lamina medullaris lateralis on the coronal AC plane. Select intersection of the most inferior tip of lamina medullaris lateralis with the most lateral tip of the anterior commissure, or substantia innominata, or nucleus centralis, when present: 13. Globus pallidus medialis was defined as the most medial point of globus pallidus medialis on the coronal AC plane. Select intersection of the most inferior tip of the capsula interna with anterior commissure: 14. Chiasma opticum was defined as the midpoint picked at the visual center of chiasma opticum; 15. Infundibulum was defined as the midpoint picked at the visual center of infundibulum; 16. Insula superior was defined as the most superior point of insula on the coronal AC plane. Select intersection of the most superior tip of insula with white matter, or the point of intersection between the most superior continuation of sulcus circularis insulae with white matter, when visible: 17. Insula inferior was defined as the most inferior point of insula on the coronal AC plane. Select intersection of the most inferior tip of insula with white matter, or the most inferior continuation of sulcus circularis insulae with white matter, when visible: 18. Heschl’s gyrus anterior was defined as the most anterior presentation of Heschl’s gyrus on coronal plane where white matter penetrates Heschl’s gyrus. Select the point at the visual center of Heschl’s gyrus (located within the white matter penetration): 19. Heschl’s gyrus posterior was defined as the most posterior presentation of Heschl’s gyrus on coronal plane where there is a visible first transverse fissure, or posterior tip of insula. Select point at the visual center of Heschl’s gyrus (located within the white matter penetration): 20. Lateral geniculate nucleus (LGN) was defined as the point picked at the visual center of LGN on the most anterior coronal plane where there is a full thickness of LGN visible; 21. Posterior commissure corresponds to the midpoint of

258

GRACHEV ET AL.

TABLE 2 Anatomic Landmarks Protocol for the Reference Brain Right hemisphere X (A) Coronal 1. Lateral ventricle anterior 2. Temporal lobe anterior 3. Isthmus of temporal lobe 4. Anterior commissure (coronal AC plane) 5. Lateral ventricle superior (coronal AC plane) 6. Lateral ventricle inferior (coronal AC plane) 7. Caudate nucleus superior (coronal AC plane) 8. Caudate nucleus inferior (coronal AC plane) 9. Putamen superior (coronal AC plane) 10. Putamen inferior (coronal AC plane) 11. Lenticulate superior notch (coronal AC plane) 12. Lenticulate inferior notch (coronal AC plane) 13. Globus pallidus medial (coronal AC plane) 14. Chiasma opticum (coronal AC plane) 15. Infundibulum (coronal AC plane) 16. Insula superior (coronal AC plane) 17. Insula inferior (coronal AC plane) 18. Heschl’s gyrus anterior 19. Heschl’s gyrus posterior 20. Lateral geniculate nucleus 21. Posterior commissure 22. Calcarine fissure anterior 23. Lateral ventricle posterior 24. Collateral fissure posterior

16 37 19 0 13 4 17 6 24 21 17 16 8 0 0 29 29 49 36 23 0 12 14 4

(B) Axial 25. Precentral fissure medial (axial AC-HM plane) 26. Precentral fissure lateral (axial AC-HM plane) 27. Central fissure medial (axial AC-HM plane) 28. Central fissure lateral (axial AC-HM plane) 29. Postcentral fissure medial (axial AC-HM plane) 30. Postcentral fissure lateral (axial AC-HM plane) 31. Superior frontal fissure anterior (axial AC-PaCiG plane) 32. Superior frontal fissure posterior (axial AC-PaCiG plane) 33. Precentral fissure medial (axial AC-PaCiG plane) 34. Precentral fissure lateral (axial AC-PaCiG plane) 35. Central fissure medial (axial AC-PaCiG plane) 36. Central fissure lateral (axial AC-PaCiG plane) 37. Postcentral fissure medial (axial AC-PaCiG plane) 38. Postcentral fissure lateral (axial AC-PaCiG plane) 39. Cingulate fissure—hemispheric margin (axial AC-PaCiG plane) 40. Postcentral fissure—intraparietal fissure (axial AC-PaCiG plane) 41. Jensen’s fissure—intraparietal fissure (axial AC-PaCiG plane) 42. Jensen’s fissure—hemispheric margin (axial AC-PaCiG plane) 43. Medial orbital sulcus anterior (axial chiasmatic plane) 44. Medial orbital sulcus posterior (axial chiasmatic plane) 45. Lateral orbital sulcus anterior (axial chiasmatic plane) 46. Lateral orbital sulcus posterior (axial chiasmatic plane) 47. Arcuate orbital sulcus medial (axial chiasmatic plane) 48. Arcuate orbital sulcus lateral (axial chiasmatic plane) 49. Olfactory fissure anterior (axial chiasmatic plane) 50. Olfactory fissure posterior (axial chiasmatic plane) 51. Lateral ventricle anterior (axial AC-PC plane) 52. Caudate nucleus anterior (axial AC-PC plane) 53. Putamen anterior (axial AC-PC plane) 54. Putamen medial (axial AC-PC plane) 55. Putamen posterior (axial AC-PC plane) 56. Lenticulate anterior notch (axial AC-PC plane) 57. Lenticulate posterior notch (axial AC-PC plane) 58. Insula anterior (axial AC-PC plane) 59. Insula posterior (axial AC-PC plane) 60. Sylvian fissure—hemispheric margin (axial AC-PC plane) 61. Lateral ventricle posterior (axial AC-PC plane) 62. Lateral ventricle anterior (axial CC plane) 63. Caudate nucleus anterior (axial CC plane) 64. Caudate nucleus posterior (axial CC plane) 65. Putamen anterior (axial CC plane)

4 25 11 32 13 31 32 30 24 43 22 49 26 51 3 28 30 40 13 18 27 30 21 41 7 11 15 17 22 14 28 15 25 29 35 47 29 17 19 6 23

y 25 23 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 218 229 224 224 239 281 2102 228 210 247 224 254 235 22 21 217 3 231 213 240 241 245 243 250 266 53 24 42 35 12 31 52 5 21 19 16 4 222 2 218 26 227 15 256 23 18 21 12

Left hemisphere z

X

y

z

4 221 223 0 26 4 23 1 16 27 8 23 1 214 219 17 211 8 10 22 0 21 3 26

14 34 25 0 13 3 17 4 24 24 17 20 9 0 0 29 35 49 37 22 0 14 19 16

25 25 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 25 225 224 224 237 277 2101

4 217 218 0 27 5 25 2 16 24 10 21 1 214 219 19 211 5 13 21 0 0 11 24

68 68 68 68 68 68 50 50 50 50 50 50 50 50 50 50 50 50 214 214 214 214 214 214 214 214 0 0 0 0 0 0 0 0 0 0 0 8 8 8 8

2 25 7 31 7 32 23 31 26 46 21 49 27 53 2 40 35 53 15 18 26 26 18 34 7 9 11 14 17 12 29 11 24 25 35 46 33 14 17 4 22

222 26 237 221 249 233 36 8 212 3 226 29 237 225 239 232 247 245 49 26 32 24 15 25 53 6 22 21 17 7 219 4 215 28 220 17 239 24 21 0 15

68 68 68 68 68 68 50 50 50 50 50 50 50 50 50 50 50 50 214 214 214 214 214 214 214 214 0 0 0 0 0 0 0 0 0 0 0 8 8 8 8

259

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

TABLE 2—Continued Right hemisphere

(B) Axial (Continued) 66. Putamen medial (axial CC plane) 67. Putamen posterior (axial CC plane) 68. Lenticulate anterior notch (axial CC plane) 69. Lenticulate posterior notch (axial CC plane) 70. Insula anterior (axial CC plane) 71. Insula posterior (axial CC plane) 72. Sylvian fissure—hemispheric margin (axial CC plane) 73. Lateral ventricle posterior (axial CC plane) (C) Midsaggital 74. Cerebrum anterior (AC-PC line) 75. Paracingulate gyrus anterior (AC-PC line) 76. Cingulate gyrus anterior (AC-PC line) 77. Cerebrum posterior (AC-PC line) 78. Corpus callosum anterior 79. Corpus callosum posterior 80. Corpus callosum posterior angle of genu 81. Corpus callosum posterior tip of genu 82. Corpus callosum fornix junction 83. Corpus callosum inferior notch of splenium 84. Corpus callosum splenium center 85. Corpus callosum posterior tip of splenium 86. Cerebrum anterior (CC line) 87. Paracingulate gyrus anterior (CC line) 88. Cingulate gyrus anterior (CC line) 89. Fornix (CC line) 90. Cerebrum posterior (CC line) 91. Thalamus center 92. Cerebrum superior (AC line) 93. Paracingulate gyrus superior (AC line) 94. Cingulate gyrus superior (AC line) 95. Corpus callosum superior (AC line) 96. Corpus callosum inferior (AC line) 97. Cerebrum superior (PC line) 98. Cingulate gyrus superior (PC line) 99. Corpus callosum superior (PC line) 100. Corpus callosum inferior (PC line) 101. Subparietal fissure—cingulate fissure 102. Subparietal fissure—calcarine fissure 103. Calcarine fissure—parietooccipital fissure 104. Parietooccipital fissure—hemisphere margin 105. Calcarine fissure—hemisphere margin 106. Cuneal fissure 1 anterior 107. Cuneal fissure 1 posterior 108. Cuneal fissure 2 anterior 109. Cuneal fissure 2 posterior (D) Lateral saggital 110. Anterior horizontal ramus of sylvian fissure anterior 111. Anterior horizontal ramus of sylvian fissure—sylvian fissure 112. Anterior ascendant ramus of sylvian fissure—sylvian fissure 113. Anterior ascendant ramus of sylvian fissure—inferior frontal fissure 114. Inferior frontal fissure—precentral fissure 115. Inferior frontal fissure—hemisphere margin 116. Precentral fissure—sylvian fissure 117. Precentral fissure—hemisphere margin 118. Central fissure—sylvian fissure 119. Central fissure—hemisphere margin 120. Postcentral fissure—sylvian fissure 121. Postcentral fissure—hemisphere margin 122. Sylvian fissure posterior 123. Posterior ascendant ramus of sylvian fissure 124. Jensen’s fissure anterior-inferior 125. Jensen’s fissure—hemisphere margin 126. Angular fissure—anterior occipital fissure 127. Superior temporal fissure anterior 128. Inferior temporal fissure anterior

X

y

16 26 16 24 34 32 55 27

4 226 2 221 28 230 27 258

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

69 50 37 2100 26 239 15 11 217 229 234 237 68 54 41 0 298 212 0 0 0 0 0 224 224 224 224 242 262 267 285 299 279 286 280 293

48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48

31 15 15 16 6 31 7 2 216 213 225 233 244 250 254 269 265 8 28

Left hemisphere z

X

y

z

8 8 8 8 8 8 8 8

15 27 15 23 30 32 54 29

5 222 3 217 30 225 3 254

8 8 8 8 8 8 8 8

0 0 0 0 7 13 7 3 22 15 14 8 7 7 7 7 7 9 68 57 40 28 20 71 46 28 20 48 19 9 33 21 16 1 21 24

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

68 50 39 2100 25 240 14 10 218 229 235 237 68 52 40 21 2100 213 0 0 0 0 0 224 224 224 224 241 263 268 287 298 280 287 282 293

0 0 0 0 7 13 7 3 22 15 13 9 7 7 7 7 7 9 69 53 41 28 20 71 45 28 20 49 20 10 34 28 15 1 22 24

2 21 21 11 27 23 1 45 13 51 15 55 19 32 14 36 10 216 237

48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48

31 15 15 20 13 29 3 1 214 29 222 221 230 236 242 246 259 11 28

5 2 2 23 30 32 10 49 16 51 19 56 21 48 29 55 16 25 234

260

GRACHEV ET AL.

FIG. 3. Localization of landmarks on the coronal AC plane: 4. Anterior Commissure; 5. Lateral ventricle superior; 6. Lateral ventricle inferior; 7. Caudate nucleus superior; 8. Caudate nucleus inferior; 9. Putamen superior; 10. Putamen inferior; 11. Lenticulate superior notch; 12. Lenticulate inferior notch; 13. Globus pallidus medialis; 14. Chiasma opticum; 15. Infundibulum; 16. Insula superior; 17. Insula inferior.

decussation of PC on the coronal plane where there is a distinct presentation of PC; 22. Calcarine fissure anterior was defined as the point of intersection between the most anterior tip of calcarina fissure and hemispheric margin; 23. Lateral ventricle posterior was defined as the point picked at the visual center of posterior horn of lateral ventricle which presented on the most caudal coronal plane; 24. Collateral fissure posterior was defined as the point of intersection between the most posterior tip of collateral fissure and hemispheric margin which presented on the most caudal coronal plane. (B) Axial landmarks. (localization of landmarks Nos. 25–30, 43–50, 62–73 presented on Figs. 4–6): 25. Precentral fissure medial was defined as the most medial point of intersection between the precentral fissure and hemispheric margin on the axial AC–HM plane; 26. Precentral fissure lateral was defined as the most lateral point of intersection between the precentral fissure and hemispheric margin on the axial

AC–HM plane; 27. Central fissure medial was defined as the point picked at the most medial tip of central fissure on the axial AC–HM plane (it approaches but does not reach the hemispheric margin); 28. Central fissure lateral was defined as the most lateral point of intersection between the central fissure and hemispheric margin on the axial AC–HM plane; 29. Postcentral fissure medial was defined as the most medial point of intersection between the postcentral fissure and hemispheric margin on the axial AC–HM plane. If it does not reach the hemispheric margin, pick the point located as far medially as visible; 30. Postcentral fissure lateral was defined as the most lateral point of intersection between the postcentral fissure and hemispheric margin on the axial AC–HM plane. If the lateral tip of postcentral fissure is thick enough to choose a point, pick the visual center of the fissure– hemispheric gap; 31. Superior frontal fissure anterior was defined as the most anterior point of intersection between the superior frontal fissure and hemispheric

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

261

FIG. 4. Localization of landmarks on the axial AC-HM plane: 25. Precentral fissure medial; 26. Precentral fissure lateral; 27. Central fissure medial; 28. Central fissure lateral; 29. Postcentral fissure medial; 30. Postcentral fissure lateral.

margin on the axial AC–PaCiG plane; 32. Superior frontal fissure posterior was defined as the point of intersection between the superior frontal fissure and precentral fissure on the axial AC–PaCiG plane. If it does not reach the precentral fissure, pick the point located as far caudally as visible; 33. Precentral fissure medial was defined as the most medial point of precentral fissure on the axial AC–PaCiG plane (it approaches but does not reach the hemispheric margin); 34. Precentral fissure lateral was defined as the most lateral point of intersection between the precentral fissure and hemispheric margin on the axial AC–PaCiG plane; 35. Central fissure medial was defined as the most medial point of central fissure on the axial AC–PaCiG plane (it approaches but does not reach the hemispheric margin); 36. Central fissure lateral was defined as the most lateral point of intersection between the central fissure and hemispheric margin on the axial AC–PaCiG plane; 37. Postcentral fissure medial was defined as the most medial point of postcentral fissure on the axial AC–

PaCiG plane (it approaches but does not reach the hemispheric margin); 38. Postcentral fissure lateral was defined as the most lateral point of intersection between the postcentral fissure and hemispheric margin on the axial AC–PaCiG plane; 39. Cingulate fissure– hemispheric margin was defined as the most medial point of intersection between the cingulate fissure and hemispheric margin on the axial AC–PaCiG plane; 40. Postcentral fissure–intraparietal fissure was defined as the most anterior point of intersection between the postcentral fissure and intraparietal fissure on the axial AC–PaCiG plane; 41. Jensen’s fissure–intraparietal fissure was defined as the point of intersection between the medial tip of Jensen’s fissure and intraparietal fissure on the axial AC–PaCiG plane; 42. Jensen’s fissure–hemispheric margin was defined as the point of intersection between the lateral tip of Jensen’s fissure and hemispheric margin on the axial AC–PaCiG plane; 43. Medial orbital sulcus anterior was defined as the point of intersection between the anterior tip of medial

262

GRACHEV ET AL.

FIG. 5. Localization of landmarks on the axial chiasmatic plane: 43. Medial orbital sulcus anterior; 44. Medial orbital sulcus posterior; 45. Lateral orbital sulcus anterior; 46. Lateral orbital sulcus posterior; 47. Arcuate orbital sulcus medial; 48. Arcuate orbital sulcus lateral; 49. Olfactory fissure anterior; 50. Olfactory fissure posterior.

orbital sulcus and hemispheric margin on the axial chiasmatic plane; 44. Medial orbital sulcus posterior was defined as the most posterior point of medial orbital sulcus on the axial chiasmatic plane (it approaches but does not reach the arcuate orbital sulcus); 45. Lateral orbital sulcus anterior was defined as the point of intersection between the anterior tip of lateral orbital sulcus and hemispheric margin on the axial chiasmatic plane; 46. Lateral orbital sulcus posterior was defined as the point of intersection between the posterior tip of lateral orbital sulcus and arcuate orbital sulcus on the axial chiasmatic plane; 47. Arcuate

orbital sulcus medial was defined as the most medial and posterior point of arcuate orbital sulcus on the axial chiasmatic plane; 48. Arcuate orbital sulcus lateral was defined as the most lateral point of arcuate orbital sulcus on the axial chiasmatic plane (if it does not reach the hemispheric margin), or the point of intersection between the lateral tip of arcuate orbital sulcus and hemispheric margin (where it is visible); 49. Olfactory fissure anterior was defined as the most anterior point of olfactory fissure on the axial chiasmatic plane (it approaches but does not reach the hemispheric margin); 50. Olfactory fissure posterior

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

263

FIG. 6. Localization of landmarks on the axial CC plane: 62. Lateral ventricle anterior; 63. Caudate nucleus anterior; 64. Caudate nucleus posterior; 65. Putamen anterior; 66. Putamen medial; 67. Putamen posterior; 68. Lenticulate anterior notch; 69. Lenticulate posterior notch; 70. Insula anterior; 71. Insula posterior; 72. Sylvian fissure-hemispheric margin; 73. Lateral ventricle posterior.

was defined as the most posterior point of olfactory fissure on the axial chiasmatic plane; 51. Lateral ventricle anterior was defined as the most anterior point of the anterior horn of lateral ventricle on the axial AC–PC plane; 52. Caudate nucleus anterior was defined as the most anterior point of the head of caudate on the axial AC–PC plane; 53. Putamen anterior was defined as the most anterior point of putamenon the axial AC–PC plane; 54. Putamen medial was defined as the most medial point of putamen on the axial AC–PC plane; 55. Putamen posterior was defined as the most posterior point of putamen on the axial AC–PC plane; 56. Lenticulate anterior notch was defined as the most anterior point of lamina medullaris lateralis (identified medially of putamen by the interface with the adjacent globus pallidus) on the axial AC–PC plane. Select intersection of anterior tip of lamina medullaris lateralis with the capsula interna;

57. Lenticulate posterior notch was defined as the most posterior point of lamina medullaris lateralis (the point of intersection between the posterior tip of lamina medullaris lateralis and white matter) on the axial AC–PC plane; 58. Insula anterior was defined as the most anterior point of insula on the axial AC–PC plane. Select intersection of the most anterior tip of insula with white matter, or the point of intersection between the most anterior continuation of sulcus circularis insulae with white matter, when visible; 59. Insula posterior was defined as the most posterior point of insula on the axial AC–PC plane. Select intersection of the most posterior tip of insula with white matter, or the point of intersection between the most posterior continuation of sulcus circularis insulae with white matter, when visible; 60. Sylvian fissure–hemispheric margin was defined as the point of intersection between the most lateral tip of Sylvian fissure and hemispheric

264

GRACHEV ET AL.

margin on the axial AC–PC plane; 61. Lateral ventricle posterior was defined as the most posterior point of posterior horn of lateral ventricle on the axial AC–PC plane; 62. Lateral ventricle anterior was defined as the most anterior point of anterior horn of lateral ventricle on the axial CC plane; 63. Caudate nucleus anterior was defined as the most anterior point of the head of caudate on the axial CC plane; 64. Caudate nucleus posterior was defined as the most posterior point of the head of caudate on the axial CC plane; 65. Putamen anterior was defined as the most anterior point of putamen on the axial CC plane; 66. Putamen medial was defined as the most medial point of putamen on the axial CC plane; 67. Putamen posterior was defined as the most posterior point of putamen on the axial CC plane; 68. Lenticulate anterior notch was defined as the most anterior point of lamina medullaris lateralis (identified medially of putamen by the interface with the adjacent globus pallidus) on the axial CC plane. Select intersection of anterior tip of lamina medullaris lateralis with the capsula interna; 69. Lenticulate posterior notch was defined as the most posterior point of lamina medullaris lateralis (the point of intersection between the posterior tip of lamina medullaris lateralis and white matter) on the axial CC plane; 70. Insula anterior was defined as the most anterior point of insula on the axial CC plane. Select intersection of the most anterior tip of insula with white matter, or the point of intersection between the most anterior continuation of sulcus circularis insulae with white matter, when visible; 71. Insula posterior was defined as the most posterior point of insula on the axial CC plane. Select intersection of the most posterior tip of insula with white matter, or the point of intersection between the most posterior continuation of sulcus circularis insulae with white matter, when visible; 72. Sylvian fissure–hemispheric margin was defined as the point of intersection between the most lateral tip of Sylvian fissure and hemispheric margin on the axial CC plane; 73. Lateral ventricle posterior was defined as the most posterior point of posterior horn of lateral ventricle on the axial CC plane. (C) Midsagittal landmarks. (localization of landmark Nos. 74–109 presented on Fig. 7): 74. Cerebrum anterior was defined as the most anterior point of intersection hemispheric margin with AC–PC line (horizontal line drawn through AC and PC points) on the midsagittal planes; 75. Paracingulate gyrus anterior was defined as the point of intersection AC–PC line with paracingulate fissure on the midsagittal planes; 76. Cingulate gyrus anterior was defined as the point of intersection AC–PC line with cingulate fissure on the midsagittal planes; 77. Cerebrum posterior was defined as the most posterior point of intersection hemispheric

margin with AC–PC line on the midsagittal planes; 78. Corpus callosum anterior was defined as the most anterior point of corpus callosum on the midsagittal planes. Select intersection of anterior tip of callosal fissure with CC line (horizontal line drawn through posterior angle of genu of corpus callosum) when the most anterior point of corpus callosum is not visible; 79. Corpus callosum posterior was defined as the most posterior point of corpus callosum on the midsagittal planes; 80. Corpus callosum posterior angle of genu corresponds to the point of origin of CC line on the midsagittal planes; 81. Corpus callosum posterior tip of genu corresponds to the location of the most posterior point of corpus callosum posterior tip of genu on the midsagittal planes; 82. Corpus callosum fornix junction was defined as the midpoint of connection between corpus fornicis and body of corpus callosum on the midsagittal planes; 83. Corpus callosum inferior notch of splenium corresponds to the location of inferior notch on the midsagittal planes; 84. Corpus callosum splenium center was defined as the point picked at the visual center of splenium on the midsagittal planes; 85. Corpus callosum posterior tip of splenium corresponds to the location of the most posterior point of corpus callosum posterior tip of splenium on the midsagittal planes; 86. Cerebrum anterior was defined as the most anterior point of intersection hemispheric margin with CC line on the midsagittal planes; 87. Paracingulate gyrus anterior was defined as the point of intersection CC line with paracingulate fissure on the midsagittal planes; 88. Cingulate gyrus anterior was defined as the point of intersection CC line with cingulate fissure on the midsagittal planes; 89. Fornix was defined as the midpoint of intersection CC line with corpus fornicis on the midsagittal planes, or the point of intersection CC horizontal line with AC vertical line; 90. Cerebrum posterior was defined as the most posterior point of intersection hemispheric margin with CC line on the midsagittal planes; 91. Thalamus center was defined as the point picked at the visual center of thalamus proper on the midsagittal planes; 92. Cerebrum superior was defined as the most superior point of intersection hemispheric margin with AC vertical line (a line drawn perpendicular to the AC–PC line through the AC point) on the midsagittal planes; 93. Paracingulate gyrus superior was defined as the point of intersection AC vertical line with paracingulate fissure on the midsagittal planes; 94. Cingulate gyrus superior was defined as the point of intersection AC vertical line with cingulate fissure on the midsagittal planes; 95. Corpus callosum superior was defined as the most superior point of intersection callosal fissure with AC vertical line on the midsagittal planes; 96. Corpus callosum inferior was defined as the most inferior point of intersection body of corpus callosum with AC vertical line on the midsagit-

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

265

FIG. 7. Localization of landmarks on the midsagittal plane: 74. Cerebrum anterior (AC-PC line); 75. Paracingulate gyrus anterior (AC-PC line); 76. Cingulate gyrus anterior (AC-PC line); 77. Cerebrum posterior (AC-PC line); 78. Corpus callosum anterior; 79. Corpus callosum posterior; 80. Corpus callosum posterior angle of genu; 81. Corpus callosum posterior tip of genu; 82. Corpus callosum fornix junction; 83. Corpus callosum inferior notch of splenium; 84. Corpus callosum splenium center; 85. Corpus callosum posterior tip of splenium; 86. Cerebrum anterior (CC line); 87. Paracingulate gyrus anterior (CC line); 88. Cingulate gyrus anterior (CC line); 89. Fornix; 90. Cerebrum posterior (CC line); 91. Thalamus center; 92. Cerebrum superior (AC vertical line); 93. Paracingulate gyrus superior (AC vertical line); 94. Cingulate gyrus superior (AC vertical line); 95. Corpus callosum superior (AC vertical line); 96. Corpus callosum inferior (AC vertical line); 97. Cerebrum superior (PC vertical line); 98. Cingulate gyrus superior (PC vertical line); 99. Corpus callosum superior (PC vertical line); 100. Corpus callosum inferior (PC vertical line); 101. Subparietal fissure-cingulate fissure; 102. Subparietal fissure-calcarine fissure; 103. Calcarine fissure-parietooccipital fissure; 104. Parietooccipital fissure-hemisphere margin; 105. Calcarine fissure-hemisphere margin; 106. Cuneal fissure 1 anterior; 107. Cuneal fissure 1 posterior; 108. Cuneal fissure 2 anterior; 109. Cuneal fissure 2 posterior.

tal planes; 97. Cerebrum superior was defined as the most superior point of intersection hemispheric margin with PC vertical line (a line drawn perpendicular to the AC–PC line through the PC point) on the midsagittal planes; 98. Cingulate gyrus superior was defined as the point of intersection PC vertical line with cingulate fissure on the midsagittal planes; 99. Corpus callosum superior was defined as the most superior point of

intersection callosal fissure with PC vertical line on the midsagittal planes; 100. Corpus callosum inferior was defined as the most inferior point of intersection body of corpus callosum with PC vertical line on the midsagittal planes; 101. Subparietal fissure–cingulate fissure was defined as the point of intersection subparietal fissure with cingulate on the midsagittal planes; 102. Subparietal fissure–calcarine fissure was defined as

266

GRACHEV ET AL.

FIG. 8. Localization of landmarks on the lateral sagittal plane: 110. Anterior horizontal ramus of Sylvian fissure anterior; 111. Anterior horizontal ramus of Sylvian fissure-Sylvian fissure; 112. Anterior ascendant ramus of Sylvian fissure-Sylvian fissure; 113. Anterior ascendant ramus of Sylvian fissure-inferior frontal fissure; 114. Inferior frontal fissure-precentral fissure; 115. Inferior frontal fissure-hemisphere margin; 116. Precentral fissure-Sylvian fissure; 117. Precentral fissure-hemisphere margin; 118. Central fissure-Sylvian fissure; 119. Central fissure-hemisphere margin; 120. Postcentral fissure-Sylvian fissure; 121. Postcentral fissure-hemisphere margin; 122. Sylvian fissure posterior; 123. Posterior ascendant ramus of Sylvian fissure; 124. Jensen’s fissure anterior-inferior; 125. Jensen’s fissure-hemisphere margin; 126. Angular fissure-anterior occipital fissure; 127. Superior temporal fissure anterior; 128. Inferior temporal fissure anterior.

the point of intersection posterior tip of subparietal fissure with calcarine on the midsagittal planes, or the most inferior point of posterior tip of subparietal fissure when it approaches but does not reach the calcarine; 103. Calcarine fissure–parietooccipital fissure was defined as the point of intersection parietooccipital fissure with calcarine on the midsagittal planes; 104. Parietooccipital fissure–hemisphere margin was defined as the point of intersection of parietooccipital fissure with hemispheric margin on the midsagittal planes; 105. Calcarine fissure–hemisphere margin was defined as the point of intersection calcarine fissure with hemi-

spheric margin on the midsagittal planes; 106. Cuneal fissure 1 anterior was defined as the most anteriorsuperior point of cuneal fissure 1 (cuneal fissure 1 distinguished by it specific topographic localization—it follows closely and parallel to the calcarine fissure) on the midsagittal planes, or the point of intersection cuneal fissure 1 with parietooccipital fissure; 107. Cuneal fissure 1 posterior was defined as the most posterior-inferior point of cuneal fissure 1 on the midsagittal planes (it approaches but does not reach the hemispheric margin); 108. Cuneal fissure 2 anterior was defined as the most anterior-superior point of

EVALUATION OF INTERSUBJECT REGISTRATION ACCURACY

cuneal fissure 2 (cuneal fissure 2 distinguished by it localization between cuneal fissure 1 and hemispheric margin, it follows closely and parallel to the cuneal fissure 1) on the midsagittal planes, or the point of intersection cuneal fissure 2 with parietooccipital fissure; 109. Cuneal fissure 2 posterior was defined as the most posterior-inferior point of cuneal fissure 2 on the midsagittal planes (it approaches but does not reach the hemispheric margin). (D) Lateral sagittal landmarks. (localization of landmark Nos. 110–128 presented on Fig. 8): 110. Anterior horizontal ramus of Sylvian fissure anterior corresponds to the most anterior visible point of anterior horizontal ramus of Sylvian fissure on the lateral sagittal planes specified above; 111. Anterior horizontal ramus of Sylvian fissure–Sylvian fissure was defined as the point origin of anterior horizontal ramus from Sylvian fissure on the lateral sagittal planes (most posterior visible point of intersection anterior horizontal ramus with Sylvian fissure or its ascending stem); 112. Anterior ascendant ramus of Sylvian fissure– Sylvian fissure was defined as the point origin of anterior ascendant ramus from Sylvian fissure on the lateral sagittal planes (most inferior visible point of intersection anterior ascendant ramus with Sylvian fissure); 113. Anterior ascendant ramus of Sylvian fissure–inferior frontal fissure was defined as the most superior point of anterior ascendant ramus of Sylvian fissure on the lateral sagittal planes (it approaches but does not reach the inferior frontal fissure); 114. Inferior frontal fissure–precentral fissure was defined as the point of intersection posterior tip of inferior frontal fissure with precentral fissure on the lateral sagittal planes; 115. Inferior frontal fissure–hemisphere margin was defined as the point of intersection anterior tip of inferior frontal fissure with hemispheric margin on the lateral sagittal planes; 116. Precentral fissure– Sylvian fissure was defined as the point of intersection of inferior tip of precentral fissure with Sylvian fissure on the lateral sagittal planes. If it approaches but does not reach the Sylvian fissure, select the most inferior point of precentral fissure; 117. Precentral fissure– hemisphere margin was defined as the point of intersection of superior tip of precentral fissure with hemispheric margin on the lateral sagittal planes; 118. Central fissure–Sylvian fissure was defined as the point of intersection of inferior tip of central fissure with Sylvian fissure on the lateral sagittal planes. If it approaches but does not reach the Sylvian fissure, select the most inferior point of central fissure; 119. Central fissure–hemisphere margin was defined as the point of intersection of superior tip of central fissure with hemispheric margin on the lateral sagittal planes; 120. Postcentral fissure–Sylvian fissure was defined as the point of intersection of inferior tip of postcentral fissure with Sylvian fissure on the lateral sagittal

267

planes. If it approaches but does not reach the Sylvian fissure, select the most inferior point of postcentral fissure; 121. Postcentral fissure–hemisphere margin was defined as the point of intersection of superior tip of postcentral fissure with hemispheric margin on the lateral sagittal planes; 122. Sylvian fissure posterior was defined as the most posterior point of Sylvian fissure on the lateral sagittal planes (the point of conjunction of Sylvian fissure with its posterior ascendant ramus); 123. Posterior ascendant ramus of Sylvian fissure was defined as the most superior-posterior point of posterior ascendant ramus on the lateral sagittal planes; 124. Jensen’s fissure anterior-inferior was defined as the most anterior-inferior point of Jensen’s intermediate fissure on the lateral sagittal planes (if there are two Jensen’s fissures, select the most distinct and prominant one which more often identified in the gap between the posterior ascendant ramus of Sylvian fissure and angular fissure); 125. Jensen’s fissure–hemisphere margin was defined as the point of intersection Jensen’s fissure with hemispheric margin on the lateral sagittal planes; 126. Angular fissure–anterior occipital fissure was defined as the point of intersection angular fissure with anterior occipital fissure on the lateral sagittal planes; 127. Superior temporal fissure anterior was defined as the point of intersection anterior tip of superior temporal fissure with hemispheric margin on the lateral sagittal planes; 128. Inferior temporal fissure anterior was defined as the point of intersection anterior tip of inferior temporal fissure with hemispheric margin on the lateral sagittal planes. REFERENCES Arndt, S., Rajarethinam, R., Cizadlo, T., O’Leary, D., Downhill, J., and Andreasen, N. C. 1996. Landmark-based registration and measurement of magnetic resonance images: A reliability study. Psychiatry Res. Neuroimaging 67:145–154. Bookstein, F. L. 1989. Principal warps: Thin-plate splines and the decomposition of deformations. IEEE Trans. Pattern Anal. Mach. Intell. 11:567–585. Bookstein, F. L. 1994. Landmarks, edges, morphometrics and the brain atlas problem. In Functional Neuroimaging: Technical Foundations (R. W. Thatcher, M. Hallett, T. Zeffiro, J. E. Roy, and M. Huerta, Eds.), pp. 107–119. Academic Press, San Diego, CA. Caviness, V. S., Meyer, J., Makris, N., and Kennedy, D. N. 1996. MRI-based topographic parcellation of human neocortex: An anatomically specified method with estimate of reliability. J. Cognitive Neurosci. 8(6):566–587. Collins, D. L., Holmes, C. J., Peters, T. M., Evans, A. C. 1995. Automatic 3-D model-based neuroanatomical segmentation. Hum. Brain Map. 3:190–208. Collins, D. L., Evans, A. C. 1997. Animal: validation and application of nonlinear registration-based segmentation. Int. J. Pattern Recog. Art. Intel. 11:1271–1294. Davatzikos, C. 1996. Spatial normalization of 3D brain images using deformable models. J. Comput. Assist. Tomogr. 20(4):656–665. Dongsheng, C., and Yuwen, L. 1993. Standards for skull-to-photo

268

GRACHEV ET AL.

superimposition. In Forensic Analysis of the Skull (M. Y. Iscan and R. P. Helmer, Eds.), pp. 183–198. Wiley, New York. Evans, A. C., Dai, W., Collins, L., Neelin, P., and Marett, S. 1991. Warping of a computerized 3D atlas to match brain image volumes for quantitative neuroanatomical and functional analysis. Image Process. 1445:236–246. Filipek, P. A., Richelme, C., Kennedy, D. N., and Caviness, V. S. 1994. The young adult human brain: An MRI-based morphometric analysis. Cerebral Cortex 4:344–360. Friston, K. J., Frith, C. D., Liddle, P. F., and Frackowiak, R. S. J. 1991. Plastic transformation of PET images. J. Comput. Assist. Tomogr. 15:634–639. Georg, R. M. 1993. Anatomical and artistic guidelines for forensic facial reconstruction. In Forensic Analysis of the Skull (M. Y. Iscan, and R. P. Helmer, Eds.), pp. 215–227. Wiley, New York. Grachev, I. D., Jenike, M. A., Baer, L., Rauch, S. L., O’Sullivan, R. L., Breiter, H. C., Kennedy, D. N., Keuthen, N. J., Savage, C. R., and Caviness, V. S. 1996. Magnetic resonance imaging study of the brain’s neocortex. NeuroImage 3(3):132. Grachev, I. D., Breiter, H. C., Rauch, S. L., Savage, C. R., Baer, L., Shera, D. M., Kennedy, D. N., Makris, N., Caviness, V. S., and Jenike, M. A. 1998. Structural abnormalities of frontal neocortex in obsessive-compulsive disorder. Arch. Gen. Psychiatry 55:181–182. Grenander, U., and Miller, M. I. 1994. Representation of knowledge in complex systems. J. R. Stat. Soc. B 56(3). Iosifescu, D. V., Shenton, M. E., Warfield, S. K., Kikinis, R., Dengler, J., Jolesz, F. A., and McCarley, R. W. 1997. An automated registration algorithm for measuring MRI subcortical brain structures. NeuroImage 6:13–25. Kikinis, R., Gleason, L., and Moriarty, T. M. 1996. Computer-assisted interactive three-dimensional planning for neurosurgical procedures. Neurosurgery 38(4):640–649. Miller, M. I., Christensen, G. E., Amit, Y., and Grenander, U. 1993. Mathematical textbook of deformable neuroanatomies. Proc. Natl. Acad. Sci. USA 90:11944–11948. Neelin, P., Crossman, J., Hawkes, D. J., Ma, Y., and Evans, A. C. 1993. Validation of an MRI/PET landmark registration method using 3D simulated PET images and point simulations. Comput. Med. Imaging Graphics 17(4–5):351–356. Novotny, V., Iscan, M. Y., and Loth, S. R. 1993. Morphologic and osteometric assessment of age, sex and race from the skull. In Forensic Analysis of the Skull (M. Y. Iscan and R. P. Helmer, Eds.), pp. 71–88. Wiley, New York. Paus, T., Tomaiuolo, F., Otaky, N., MacDonald, D., Petrides, M., Atlas, J., Morris, R., and Evans, A. C. 1996a. Human cingulate and paracingulate sulci: Pattern, variability, asymmetry, and probabilistic map. Cereb. Cortex 6(2):207–214. Paus, T., Otaky, N., Caramanos, Z., MacDonald, D., Zijdenbos, A., D’Avirro, D., Gutmans, D., Holmes, C., Tomaiuolo, F., and Evans, A. C. 1996b. In vivo morphometry of the intrasulcal gray matter in the human cingulate, paracingulate, and superior-rostral sulci:

Hemispheric asymmetries, gender differences and probability maps. J. Comp. Neurol. 376(4):664–673. Rademacher, J., Galaburda, A. M., Kennedy, D. N., Filipek, P. A., Caviness, V. S. Jr. 1992. Human cerebral cortex: Localization, parcellation, and morphometry with magnetic resonance imaging. J. Cogn. Neurosci. 4:352–374. Rademacher, J., Caviness, V. S., Steinmetz, H., Galaburda, A. M. 1993. Topographical variation of the human primary cortices: Implications for neuroimaging, brain mapping, and neurobiology. Cereb. Cortex. 3:313–329. Rajkowska, G., and Goldman-Rakic, P. 1995. Cytoarchitectonic definition of pre-frontal areas in the normal human cortex. II. Variability in locations of areas 9 and 46 and relationship to the Talairach coordinate system. Cereb. Cortex. 5:323–337. Schormann, T., Henn, S., and Zilles, K. 1996. A new approach to fast elastic alignment with application to human brains. Lecture Notes Comput. Sci. 1131:437–442. Talairach, J., and Tournoux, P. 1988. Co-planar Stereotaxic Atlas of the Human Brain. 3-Dimensional Proportional System: An Approach to Cerebral Imaging. Thieme Medical, New York. Thompson, P. M., Schwartz, C., Lin, R. T., Khan, A. A., and Toga, A. W. 1996a. 3D statistical analysis of sulcal variability in the human brain. J. Neurosci. 16(13):4261–4274. Thompson, P. M., and Toga, A. W. 1996b. A surface-based technique for warping three-dimensional images of the brain. IEEE Trans. Med. Imaging. 15(4):402–417. Thompson, P. M., MacDonald, D., Mega, M. S., Holmes, C. J., and Evans, A. C. 1997. Detection and mapping of abnormal brain structure with a probabilistic atlas of cortical surfaces. J. Comput. Assist. Tomogr. 21(4):567–581. Toga, A. W. 1994. Visualization and warping of multi-modality brain imagery. In Functional Neuroimaging: Technical Foundations (R. W. Thatcher, M. Hallett, T. Zeffiro, J. E. Roy, and M. Huerta, Eds.), pp. 171–180. Academic Press, San Diego, CA. Warfield, S., Dengler, J., Zaers, J., Guttmann, C. R. G., Wells, W. M., Ettinger, G. J., Hiller, J., and Kikinis, R. 1996. Automatic identification of gray matter structures from MRI to improve the segmentation of white matter lesions. J. Image Guided Surg. 1(6):326–338. Wible, C. G., Shenton, M. E., Fischer, I. A., Allard, J. E., Kikinis, R., Jolesz, F. A., Iosifescu, D. V., and McCarley, R. W. 1997. Parcellation of the human prefrontal cortex using MRI. Psychiatry Res. Neuroimaging 76:29–40. Woods, R. P., Grafton, S. T., Holmes, C. J., Cherry, S. R., and Mazziotta, J. C. 1998a. Automated image registration: I. General methods and intrasubject, intramodality validation. J. Comput. Assist. Tomogr. 22(1):139–152. Woods, R. P., Grafton, S. T., Watson, J. D., Sicotte, N. L., and Mazziotta, J. C. 1998b. Automated image registration: II. Intersubject validation of linear and nonlinear models. J. Comput. Assist. Tomogr. 22(1):153–165.