Application of an automated parcellation method to the analysis of pediatric brain volumes

PSYCHIATRY RESEARCH ELSEVIER

NEUROIMAGING

Psychiatry Research: Neuroimaging Section 76 (1997) 15-27

Application of an automated parcellation method to the analysis of pediatric brain volumes Desmond M. Kaplan a'b, Alex M.C. Liu d, Michael T. Abrams a'b, liana S. Warsofsky a, Wendy R. Kates a'b, Christopher D. White d, Walter E. Kaufmann a'b'c,Allan L. Reiss d,* aNeuroimagingLaboratory, Kennedy KriegerInstitute, 707 North Broadway, Baltimore, MD 21205, USA bDepartrnentof Psychiatry and Behavioral Sciences, Johns Hopkins UniversitySchool of Medicine, Baltimore, MD 21205, USA CDepartmentof Pathology, Neurologyand Pediatrics, Johns Hopkins UniversitySchool of Medicine, Baltimore, MD 21205, USA dDivision of Child and Adolescent Psychiatryand Child Development, Department of Psychiatry, Stanford UniversitySchool of Medicine, 401 QuarryRoad, MC 544, Stanford, CA 94305, USA Received 25 April 1997; revised 25 August 1997; accepted 18 September 1997

Abstract

New techniques in quantitative imaging are needed to accelerate understanding of brain development and function in children. In this study we evaluate the reliability and validity of an automated parcellation method for the measurement of large and small brain regions in normal and developmentally disabled children. We utilized an adaptation of the Talairach atlas to semi-automatically quantify brain volumes from 10 children with fragile X syndrome, 10 age- and gender-matched controls and 10 adult controls comparing them to 'gold standard' manually delineated regions. Excellent sensitivity, specificity, intra-class correlation and positive predictive value were achieved for large structures although results were less satisfactory for smaller structures, illustrating the limits of resolution of the method. Statistically significant differences in regional brain volumes were shown between males and females, children and adults, and individuals with fragile X and matched controls. This study demonstrates an automated method which rapidly and accurately quantifies large neuroanatomical structures, but not smaller structures. This method is sufficiently accurate to demonstrate some known anatomical differences in individuals with fragile X; the results suggest that this method could be applied to the assessment of brain volume in other neurodevelopmental disabilities. © 1997 Elsevier Science Ireland Ltd Keywords: Magnetic resonance imaging; Morphometric analysis; Talairach atlas

* Corresponding author. Tel.: + 1 415 7235446; e-mail: [email protected] 0925-4927/97/$17.00 © 1997 Elsevier Science Ireland Ltd. All rights reserved. PII S0925-4927(97) 00055-3

16

D.M. Kaplan et al. / PsychiatryResearch: Neuroimaging 76 (1997) 15-27

1. Introduction

The ability of in vivo quantitative neuroimaging studies to reveal insights into brain-behavior associations relevant to human disease has been demonstrated in a number of neurologic, neuropsychiatric and neurodevelopmental conditions such as Huntington's disease, schizophrenia, Tourette's syndrome, fragile X syndrome and trisomy 21 (Jernigan et al., 1993; Singer et al., 1993; Andreasen et al., 1994; Aylward et al., 1994; Reiss et al., 1994). Most quantitative imaging studies of the human brain utilize parcellation methods to provide more precise analysis of specific structure-function correlations. Parcellation refers to the process of subdividing the brain into (presumed) neurofunctionally relevant regions, thus permitting direct comparison of anatomy or function in one subject, or group, to another. Some parcellation methods utilize a standardized stereotaxic atlas with specific neuroanatomical anchor points and landmarks as the basis for neuroanatomical subdivision (Talairach and Tournoux, 1988; Zipursky et al., 1992). Alternatively, other parcellation methods use sulcal/gyral landmarks identified on the cerebral surface (Rademacher et al., 1992) as the basis for subdividing the brain. The latter approach relies on manual identification and delineation of specific sulcal landmarks as visualized in 2D or multiplanar images, and surface renderings (Rademacher et al., 1992; Andreasen et al., 1994, 1996). Although sulcal-based parcellation methods may provide more accurate functional subdivision of the brain, reliance on manual methods in quantitative imaging studies potentially can be quite time-consuming and require high-level neuroanatomical knowledge, thereby limiting the number of subjects that can be analyzed and the statistical power of a comparative study. Manual methods also can lead to decreased consistency and accuracy in measurement (Andreasen et al., 1994a, 1996). This is particularly true with respect to attempts to delineate and quantify complex cerebral regions defined by cortical topography known to demonstrate considerable morphologic variability in sulcal/gyral anatomy (Andreasen et al., 1994b). Accordingly, there is a need for more

automated methods to improve time efficiency and reliability of measurement so that larger subject populations can be studied more rapidly and accurately. This need is especially evident for studies of children with abnormalities of development and learning where, in relation to adults, there are only limited data on normal and abnormal brain development in a small number of subjects. Andreasen et al. (1994a, 1996), have proposed that a computer software implementation of the stereotaxic coordinate system originally proposed by Talairach and Tournoux (1988), could be used to automate the process of cerebral lobe parcellation and measurement in images derived from magnetic resonance (MR) imaging. Superimposition of the Talairach atlas on a given brain positionally normalizes cerebral volume by mapping a proportional grid system around specific neuroanatomical anchor points. With this method, the brain is divided into 1232 three-dimensional rectangular sectors that theoretically map to the same neuroanatomical regions across subjects (Talairach and Tournoux, 1988). Andreasen et al. (1996) showed that this automated atlas-based system of cerebral parcellation could be used reliably and accurately to define major regions (lobes) of the adult brain when compared to significantly more time-consuming manual tracing of these same regions through the identification of sulcal boundaries (Andreasen et al., 1994a, 1996). These studies also demonstrated the initial validity of this parcellation procedure by showing that quantitative neuroanatomical differences between adults with schizophrenia and matched controls could be demonstrated with Talairachdefined cortical regions as well as manual procedures. In the present study we attempt to investigate the applicability and validity of Talairach-defined brain regions in MR images obtained from a pediatric population consisting of normal male and female children, and children with the fragile X syndrome ranging in age from 4 to 15 years. Potential differences in the applicability of this methodology to children vs. adults are also addressed by including MR brain images derived from normal adults. Finally, we extend the

D.M. Kaplan et al. / PsychiatryResearch:Neuroimaging76 (1997)15-27 methodology described by Andreasen et al. (1996), by modifying the proposed Talairach lobe definitions and by combining these Talairach sector definitions with tissue segmentation techniques to isolate and measure specific subcortical gray nuclei. The utilization of more rapid, reliable and accurate techniques for analyzing pediatric brain morphology in quantitative imaging studies will help to accelerate our understanding of normal and abnormal brain development and function in children. 2. Method 2.1. Subjects Scans were obtained from 10 children with fragile X syndrome (five male and five female), 10 pediatric controls individually matched to fragile X subjects for gender and age, and 10 adult controls. The children with fragile X are part of an ongoing study of this disorder at our institution. Their diagnosis was confirmed using a direct DNA test for this genetic disorder (Rousseau et al., 1991). The control groups consisted of healthy volunteers, many of whom were the non-affected parents or siblings of developmentally disabled subjects. The control subjects had no history of psychiatric, neurological, developmental or learning disorders. IQ scores were assessed using standardized instruments (Wechsler, 1981, 1991; Thorndike et al., 1986). The fragile X subjects had a mean age of 9 _+ 3 years and a mean IQ of 67 + 21. The pediatric controls had a mean age of 8.5 + 2.3 years and a mean IQ of 107 + 13 while adult controls had a mean age of 36 + 7 years and a mean IQ of 104 + 15. For the preliminary study of the frontal lobe, 15 control subjects were selected (10 subjects overlapped with the previous groups). Among the 15, five were male children (age 8.47 + 1.50), five female children (9.12+ 2.66) and five adults (34.00 + 6.97). 2.2. Image acquisition and processing Coronal T1 weighted SPGR brain scans were acquired on a 1.5-T GE Signa scanner using the

17

following pulse sequence: TR -- 35-45 ms, TE = 6-7 ms, flip angle = 45°, single excitation, field of view = 20-24 cm; image matrix 256 × 128 pixels. This sequence yielded 124 Tl-weighted contiguous 1.5-mm slices covering the entire brain. Imaging data from the MRI studies were transferred to our laboratory and were processed on PowerPC workstations running the MacOS 7.5x and the software program Brainlmage (Reiss, 1997). The importing procedure converts each brain slice into 8-bit images and corrects shading artifacts across slices. Removal of non-brain material was achieved in several automated steps developed in our laboratory which included an erosion-dilation edge-detection routine to remove the skull and non-brain tissues, and a region growing algorithm to remove intra-brain vasculature (Reiss, 1997). The data were then resampled to create isotropic voxels using a Catmull-Rom spline algorithm (Browne and Gaydecki, 1987). 2.3. Manual delineation of regions of interest (ROIs) The volumes of two large structures (cerebrum, cerebellum), two medium sized structures (frontal lobe and brain stem) and two smaller regions (lateral ventricle and caudate nucleus) were measured in this initial study. Prior to the measurement of these structures, cerebrospinal fluid (CSF) was removed from the dataset (apart from the lateral ventricles). This was accomplished using a histogram-based semi-automated segmentation technique (Otsu, 1979) to categorically identify all voxels with CSF-like intensity. Non-CSF voxels were further subdivided into gray or white matter. The cerebrum (whole brain minus cerebellum and brainstem), cerebellum and brainstem were then manually circumscribed slice-by-slice on axial sections through the brain using previously established rules for measuring these regions from MRI scans (Aylward and Reiss, 1991; Reiss et al., 1995). The caudate nucleus was manually circumscribed in the coronal plane anterior to posterior until the structure disappeared from its position adjacent to the lateral ventricle. A ventricular region was obtained by placing regions of interest (ROIs) around CSF spaces to include the lateral

18

D.M. Kaplanet al. / PsychiatryResearch:Neuroimaging76 (1997)15-27

ventricles, third ventricle and the velum interpositum. The demarcation of these spaces was initiated using a region growing algorithm. The ROI was then manually detailed to ensure that only CSF in the above-mentioned spaces was included. Inter- and intra-rater reliabilities for manually delineated regions were established on five of the total sample of 30 scans. The manually delineated regions comprised the 'gold-standard' measure to which the Talairach measurement was compared.

sured using segmented images so that only pixels with CSF-like intensity were captured. Since subcortical gray matter is of lighter gray-scale intensity than cortical gray matter, a separate segmentation step for the subcortical region was performed as described in Reiss et al. (1995). This resulted in subcortical voxels being categorized as either white or gray. Caudate volume could then be measured by automatically counting all the subcortical gray-classified pixels included within the Talairach-defined ROI.

2.4. Talairach-defined regions 2. 5. Preliminary frontal lobe study

The method described by Andreasen et al. (1996) was used to delineate the Talairach-defined regions. To initiate this semi-automated procedure, the rater first identifies three anchor points (anterior commissure, posterior commissure and a midsagittal point above the axis created by the first two points). The six faces of the brain 'cube' are then automatically located by the software and a proportional stereotaxic grid is overlaid onto the brain by a linear interpolation of areas anterior to AC, posterior to PC and between AC and PC, respectively (Fig. 1). This grid is composed of 1232 three-dimensional rectangular sectors (Talairach and Tournoux, 1988; Andreasen et al., 1996). Sets of individual sectors are then grouped together to correspond to the neuroanatomicai structures of interest. Two sets of Talairach-derived ROIs were used to automatically measure the ventricles, cerebellum and cerebrum; one set of ROIs delineated the sectors originally assigned to each structure by Andreasen et al. (referred to as the 'original' Talairach definition); the other set represented our revisions of those regional definitions (referred to as the 'revised' Talairach definition). These revisions consisted of six pairs of Talairach sectors originally assigned to the cerebellum which we reassigned to the occipital and temporal lobes of the cerebrum. Automated Talairach defined volume measurements for cerebrum, cerebellum and brainstem comprised all tissue (gray and white) in these regions. Brain images were measured with non-ventricular CSF and non-brain material removed. Ventricular volume was mea-

Gold-standard ROIs were obtained for the frontal lobe of each subject by identifying the central sulcus on a multi-planar and surface rendering module of BrainImage. Accurate identification of the central sulcus was confirmed for each subject by an expert neuroanatomist/neuropathologist (author W.E.K.). The sectors for the frontal lobe were chosen if the majority of the Talairach sector was occupied by the lobe. The volume of frontal lobe tissue in each Talairach sector was then measured. 2.6. Data anatyses

Inter- and intra-rater reliabilities were calculated with the intraclass correlation coefficient, a statistical test that takes pair-wise absolute volume differences into account as well as the slope of the line corresponding to pair-wise rater agreement. In addition, the percent spatial agreement between raters was calculated as the ratio of overlapping areas to the union of areas. Groupwise differences between the mean volumes of particular brain regions were computed with analysis of variance (ANOVA). Sensitivity, specificity and positive predictive values of the Talairach-defined regions in relation to the 'gold-standard' regions were derived as demonstrated in Fig. 2. A one-tailed t-test was utilized in analyzing the volumes of caudate nuclei as a directional hypothesis had been predicted a priori (Reiss et al., 1995).

D.M. Kaplanet al. / PsychiatryResearch:Neuroimaging76 (1997)15-27

19

Fig. 1. A tri-planar view of the Talairach stereotaxic grid overlaid on a brain. A rendered three-dimensional view of the planes is displayed in the lower left. The grid divides the brain into 1232 three-dimensional rectangular sectors.

3. Results 3.1. Inter- and intra-rater reliability of gold-standard ROIs Inter-rater and intra-rater reliabilities were ob-

tained (n = 5) for all structures (i.e. cerebrum, cerebellum, brainstem, caudate gray and ventricles). T h e inter-rater reliability ranged f r o m 0.988 to 1.000 for all structures except for brainstem volume (0.765), which was lower due to the ambiguous b o u n d a r y between the midbrain and the

20

D.M. Kaplan et al. / Psychiatry Research: Neuroimaging 76 (1997) 15-27 Total Reference Region

Gold Standard Region

TN Talairach Region

Fig. 2. Sensitivity, specificity and positive predictive value. TP = true positive (gray with lines); FP = false positive (gray only); FN = false negative (lines only); TN = true negative (white only). Sensitivity is the portion of the gold-standard region correctly included within the Talairach region (sensitivity = TP/(TP + FN) × 100%), Specificity is the portion of the (non-gold standard) reference region correctly excluded from the Talairach region (Specificity = TN/(TN + FP) × 100%). Positive predictive value (PPV) is the portion of the Talairach region that consists of the gold standard region [PPV = TP/(TP + FP) × 100%].

cortex. The intra-rater reliabilities ranged from 0.925 to 1.000. Percent spatial agreement was obtained for all measurements. The values ranged from 83.6% to 99.6% agreement with standard deviations ranging from 0.008 to 0.028.

3.2. Sensitivity, specificity, intra-class correlation and positive predictive value Table 1 shows the results of sensitivity, specificity, intra-class correlations and positive predictive values for Talairach-based measurements of the neuroanatomical regions of interest. Comparison of sensitivity and specificity among the three groups of subjects revealed highly consistent values. Therefore the results presented below refer to data across all 30 subjects (the last four columns in Table 1). Our revision to the original Talairach definitions consisted of moving six pairs of sectors (left and right) originally assigned to the cerebellum, to the cerebrum (occipital and temporal lobes, specifically). Therefore compared to the

original definitions, the revised Talairach cerebral region tended to increase sensitivity slightly at the expense of specificity. The converse situation was observed for the cerebellar region. The sensitivities and specificities of the cerebellar region derived in this study (both original and revised Talairach definitions) were comparable to the values previously published by Andreasen et al. (0.90 and 0.99, respectively). Although the revised Talairach measure of cerebrum and cerebellum resulted in minimal changes in overall sensitivity and specificity compared to the original definition, the intra-class correlation for the revised cerebellum measure (corresponding to the agreement between the gold-standard and Talairach regions) was notably increased (from 0.22 to 0.79), indicating that this measure was more consistent with, and closer in volume to the gold-standard measurement (see Table 2). Although the ICC for cerebellum is lower than that of the ventricles and cerebrum, we still believe it to be an acceptable value.

~Z Table 1 Sensitivity, specificity, positive predictive value (PPV) and intra-class correlation (ICC) for Talairach-based volume measurements as compared to gold standard (exact) region of interest

E" Adult (n = 10) Sen.

Structures Cerebrum

Spec.

Control child (n = 10)

Fragile X (n = 10)

All subjects (n = 30)

ICC

Sen.

Spec.

ICC

Sen.

Spec.

ICC

Sen.

Spec.

PPV

ICC

~

,~"

Original Revised

0.98 + 0.01 0.93 + 0.04 0.98 + 0.01 0.91 + 0.04

0.99 0.99

0.97 + 0.02 0.98 _+ 0.01

0.92 + 0.03 0.88 + 0.02

0.99 0.99

0.97 ± 0.01 0.98 + 0.00

0.93 + 0.03 0.92 + 0.03

0.99 0.99

0.97 + 0.01 0.98 + 0.01

0.95 + 0.02 0.94 + 0.02

99.29 + 0.30 99.12 + 0.32

0.99 0.99

Cerebellum

Original Revised

0.90 + 0.04 0.88 + 0.03

0.97 + 0.01 0.99 + 0.01

0.28 0.81

0.90 + 0.03 0.86 + 0.03

0.97 + 0.01 0.99_+ 0.01

0.44 0.84

0.89 + 0.03 0.87 + 0.03

0.97 + 0.01 0.99 + 0.00

-0.65

0.90 + 0.03 0.87 + 0.03

0.97 + 0.03 0.99 + 0.00

78.83 + 4.87 90.40 ± 3.50

0.22 0.79

Ventricles

Original Revised

0.94 + 0.02 0.90 + 0.03

0.99 _+ 0.00 1.00 + 0.00

0.67 1.00

0.91 + 0.02 0.87 + 0.02

1.00 ± 0.00 1.00 + 0.00

0.82 0.99

0.94 + 0.03 0.88 + 0.04

1.00 _+ 0.00 1.00 + 0.00

0.94 0.97

0.94 + 0.00 0.88 + 0.04

1.00 + 0.00 1.00 + 0.00

74.82 ___10.57 91.37 + 4.95

0.84 0.98

Brainstem

0.80 + 0.02

0.99 + 0.00

--

0.84 + 0.05 0.99 + 0.00

--

0.82 + 0.02 0.99 + 0.00

--

0.81 + 0.04 0.99 + 0.00

60.32 + 6.76

--

i

Caudate

0.64 + 0.05

0.99_+ 0.00

0.41

0.69 ± 0.04

--

0.63 _+ 0.09 0.99 + 0.00

0.51

0.63 + 0.08

61.02 + 6.17

0.44

~

0.99 +_ 0.00

0.99 + 0.00

The original Talairach sectors were defined by Andreasen et al. and the revised sector definition, included minor modifications as determined in our laboratory. ICC values refer to correlations between the exact measures and the Talairach measures. The intra-class correlation for brainstem was null due to the significant volume differences between the Talairach and gold-standard measures.

•

D.M. Kaplanet al. / PsychiatryResearch:Neuroimaging76 (1997) 15-27

22

Table 2 Volume measurements from region of interests (ROI) of the gold standard (exact), original and revised Talairach (all in cc) Structures

Fragile X (n = 10)

Control child (n = 10)

Adult (n = 10)

Male (n = 15)

1201.23 + 95.12 1184.67+ 121.20 1120.55 + 107.30 1249.10+ 77.96 1173.08+ 101.84 1159.10+ 119.89 1101.17+ 112.28 1226.86+ 81.15 1186.53+ 94.55 1174.97+ 115.68 1113.71 + 111.73 1237.14+ 78.01

Female (n = 15)

Cerebrum

Gold standard Original Talairach Revised Talairach

Cerebellum

Gold standard Original Talairach Revised Talairach

141.91 + 9.95 166.21+ 13.49 141.02+ 10.20

144.07 + 14.45 164.13+ 25.12 139.31+ 20.13

137.16+ 14.87 155.84+ 14.87 133.95+ 12.85

148.57+ 11.86 172.19+ 17.19 146.46+ 13.67

133.53+ 9.80 151.93+ 13.87 129.73_ 10.77

Ventricles

Gold standard Original Talairach Revised Talairach

15.92 + 9.84 19.06 + 9.59 14.79 + 7.97

10.69 + 6.86 15.06 ± 7.46 10.45 + 5.96

17.19 + 7.11 23.46 + 7.70 17.01 + 6.76

17.18 + 9.09 22.24 + 9.27 16,44 + 7.57

12.02 + 6.72 16.15 + 7.24 11.73 + 6.32

Brainstem

Gold standard Talairach

29.06 + 3.33 42.20 + 5.91

27.63 + 3.45 40.09 + 6.59

33.39 + 3.06 40.52 + 3.64

31.98 + 3.52 43.91 + 4.13

28.07 + 3.62 37.96 + 4.97

Caudate

Gold standard Talairach

12.94 + 1.97 13.29 + 1.04

10.62 + 1.34 13.20 + 1.15

9.96 + 1.33 10.84 + 1.22

11.80 + 2.08 12.62 + 1.55

10.55 + 1.76 12.26 + 1.67

M o r e o v e r , the positive p r e d i c t i v e values also inc r e a s e d f r o m 78.83% to 90.40%, showing t h a t t h e revised Talairach region contains a higher portion o f t h e t a r g e t s t r u c t u r e (i.e. c e r e b e l l u m ) t h a n t h e original one. T h e sensitivity o f t h e original T a l a i r a c h m e a s u r e m e n t s for v e n t r i c u l a r v o l u m e was slightly h i g h e r t h a n t h e r e v i s e d m e a s u r e while t h e specificities o f the two m e a s u r e m e n t s w e r e similar. H o w e v e r , similar to t h e revised c e r e b e l l u m m e a sure, t h e intra-class c o r r e l a t i o n s a n d positive p r e dictive values for t h e r e v i s e d T a l a i r a c h m e a s u r e m e n t s w e r e a p p r e c i a b l y h i g h e r t h a n that o b s e r v e d for t h e original m e a s u r e , reflecting m o r e a c c u r a t e v o l u m e m e a s u r e m e n t a n d spatial a g r e e m e n t of t h e l a t e r a l ventricle. R e s u l t s for t h e b r a i n s t e m m e a s u r e s h o w e d a high specificity (0.99) a n d a m o d e r a t e sensitivity (0.80-0.84). Previous p u b l i c a t i o n o f T a l a i r a c h m e a s u r e m e n t s o f b r a i n s t e m w e r e n o t available f r o m t h e A n d r e a s e n et al. m a n u s c r i p t , a n d t h e r e w e r e no revisions to t h e original s e c t o r assignm e n t for this structure. T h e intra-class c o r r e l a t i o n for this s t r u c t u r e was null d u e to the significant volume differences between the Talairach a n d g o l d - s t a n d a r d m e a s u r e s . T h e positive p r e d i c tive value o f t h e b r a i n s t e m was 60.3%.

1088.54 + 73.26 1062.04+ 70.37 1079.68 + 72.08

T h e T a l a i r a c h definition for c a u d a t e was develo p e d in o u r l a b o r a t o r y a n d thus t h e r e was no p r e v i o u s definition available for c o m p a r i s o n . T h e specificity for t h e T a l a i r a c h - b a s e d c a u d a t e m e a sure was high (0.99), while the sensitivity was l o w e r (0.63-0.69). Intra-class c o r r e l a t i o n a n d positive p r e d i c t i v e value w e r e also lower t h a n t h o s e o f o t h e r structures (0.44 a n d 61.0%, r e s p e c tively).

3.3. Initial validation of Talairach assignments T a b l e 2 shows the a c t u a l v o l u m e t r i c m e a s u r e m e n t s for t h e g o l d - s t a n d a r d r e g i o n o f i n t e r e s t a n d t h e T a l a i r a c h - d e f i n e d regions. T h e results are p r e s e n t e d a c c o r d i n g to diagnosis, age a n d gender.

3.4. Direct paired comparisons of Talairach and gold-standard measures T h e accuracy o f t h e T a l a i r a c h - b a s e d m e a s u r e s was first e v a l u a t e d by p e r f o r m i n g p a i r e d t-tests c o m p a r i n g t h e g o l d - s t a n d a r d v o l u m e s to the Talairach v o l u m e s for all 30 subjects. T h e s e results s h o w e d t h a t t h e r e w e r e systematic differences b e t w e e n the g o l d - s t a n d a r d a n d T a l a i r a c h m e a sures for all structures ( P < 0.05) except for the revised T a l a i r a c h m e a s u r e for c e r e b e l l u m (t =

D.M. Kaplan et al. / PsychiatryResearch: Neuroimaging 76 (1997) 15-27

- 1.863, P = 0.073). For the cerebrum and ventricles there was a small systematic increase in the gold-standard measure compared to the Talairach measures. The opposite was observed for caudate and brainstem measures. These analyses were repeated for the group of 20 children only. Similar results were observed compared with that derived from the entire subject group (results not shown).

3.5. Disease group (fragile X) vs. controls Our previous studies showed that caudate and ventricular volumes were increased in individuals with fragile X compared to matched controls (Reiss et al., 1995). These findings were confirmed or supported in this small sample of children with fragile X (n = 10) and gender- and age-matched controls (n = 10) utilizing the gold-standard region measures (fragile X > controls: caudate t = 3.08, P = 0.003; ventricle t = 1.38, P = 0.09). The revised Talairach measures were also utilized to assess group differences for these brain regions. The revised Talairach ventricular measure produced results for group differences between fragile X and controls that were highly similar to that observed for the gold-standard volume (fragile X > controls, t = 1.38, P = 0.09). However, this was not observed for the Talairach caudate measure (t = 0.19, P = 0.43).

3.6. Male vs. female Gender differences in cerebral volume (t = 5.81, P < 0 . 0 0 0 1 ) , ventricular volume ( t = 1.77, P = 0.088), cerebellar volume (t = 3.72, P = 0.001) and brainstem (t = 3.00, P = 0.057) were observed for the gold-standard measures with all results showing greater volume in males (n = 30). Assessment of Talairach-defined regions showed that similar results were observed utilizing both original and revised definitions for c e r e b r u m (toriginal = 5.74, trevise d = 5.94, both P < 0.0001), cerebellum (toriginal = 3.55, eoriginal = 0.0014, trevise d = 3.27, Prevised = 0.0009) and brainstem (t = 3.56, P = 0.0013). Talairach-defined measurement of the ventricles also showed similar t r e n d s (toriginal = 2.00,

Poriginal = 0.055,

trevise d = 1.85,

Prevised ~---

23

0.0752). For the caudate measurement, a male > female trend was found in the gold-standard measures (t = 1.78, P = 0.0857) but not the Talairach-defined measures (t = 0.61, P = 0.55).

3. 7. Adults vs. children The gold-standard brainstem measure in adults (n = 10) was found to be significantly (t = - 3 . 9 5 , P = 0.009) larger than that of the normal children (n = 10). A similar trend was also found in the gold-standard measures of ventricles (t = - 2 . 0 8 , P = 0.052). The Talairach measurements were consistent with the gold-standard measures for the lateral v e n t r i c l e (toriginal = - 2 . 4 8 , eoriginal = 0.0233; trevise d = - 2 . 3 0 , erevised = 0.0336), but not for the brainstem (t = - 0 . 1 8 , P = 0.8606).

3.8. Preliminary frontal lobe data For the frontal lobe measurement, the sensitivities were: 0.9004 (control male children, n = 5 for all sub-groups), 0.9124 (control female children) and 0.8838 (control adults); the specificities were: 0.9754 (control male children), 0.9724 (control female children), and 0.9820 (control adults); and the positive predictive values were: 0.9494 (control male children), 0.9426 (control female children) and 0.9620 (control adults). The Talairach volume measurements for frontal lobe were (in cc): 465.56 _+_40.78 (control male children), 390.66 + 36.47 (control female children) and 416.70 + 53.75 (control adults). The comparable gold-standard volumes were, respectively (in cc): 491.18 + 45.14 (5.22% difference for control male children), 404.06+42.85 (3.32% difference for control female children) and 453.92 + 59.67 (8.20% difference for control adults). 4. Discussion

The original results reported by Andreasen et al. suggested that an automated Talairach atlasbased parcellation of the adult human brain could be used to produce acceptable estimates of the volumes of large anatomical structures such as the major cerebral lobes, cerebellum and ventricles. In the study presented here, we have par-

24

D.M. Kaplan et al. / Psychiatry Research: Neuroimaging 76 (1997) 15-27

tially replicated the findings of Andreasen et al. (1996), in a pediatric population as well as a small group of adults, albeit with modifications to the original Talairach sector definitions, particularly as they affect the smaller structures. (Andreasen had suggested modifications for smaller brain structures.) This replication is important for pediatric populations where one can not assume that methods developed for adults directly apply. We believe that our results justify the reallocations made to a few of the Andreasen (Talairach) sector definitions. While sensitivities (the ability to include relevant structures) and specificities (the ability to exclude irrelevant structures) were similar to those reported by Andreasen (0.87-0.98), significantly improved intra-class correlations (representing agreement with gold-standard volume) for cerebrum, cerebellum and lateral ventricles (0.84-0.9) were achieved. We found the atlas-based parcellation method to apply equally well across both pediatric and adult subject samples, and to a group of individuals with fragile X syndrome, a genetic condition known to be associated with abnormal morphology of the brain (Reiss et al., 1994, 1995). Utilizing this methodology, we also were able to show statistically significant differences in regional brain volumes between males and females, children and adults (bearing in mind that by age 3-4 years over 90% of adult brain weight is achieved), and individuals with fragile X syndrome and matched controls. For large brain structures (cerebellum and lateral ventricles) we achieved excellent sensitivity, specificity, intraclass correlation (an index of volumetric consistency) and positive predictive value (an index of spatial accuracy). For smaller structures (the caudate nucleus and the brainstem) lower sensitivity, specificity and intraclass correlation were obtained. The latter comparison was noteworthy as it illustrated the limits of resolution of this method; true differences in ventricular volume, but not caudate volume were detected by the atlas-based parcellation technique. Increasing the accuracy of volumetric quantification of these smaller structures might require a further subdivision of Talairach sectors thereby increasing the 'resolution' of the

atlas (although we should point out that size might be only one factor affecting the accuracy of parcellation). Our preliminary results for the frontal lobe indicate high sensitivity, specificity and positive predictive value across each of the three groups measured. Continued analyses of the other cerebral lobes and larger numbers of subjects (including children with known morphological brain abnormalities) are ongoing at our imaging laboratory. The implications of these findings are particularly important to research investigating brain development and morphology in children. Automated atlas-based parcellation techniques provide investigators with the potential to obtain volumetric results from large data sets rapidly and efficiently. The process of implementing the procedure described in this article requires approximately 10 min of computer interactive time for identification of the anterior and posterior commissures, and approximately 90 min of computer run time on a 604e-based PowerPC processor (as opposed to approx. 7 h if drawn and measured manually on the same machine). Although at this time, the atlas-based parcellation technique is not refined enough to measure small brain subregions, the data from the present study suggest that this approach can be used to obtain a reasonable estimate of major brain regions in children as represented by volumetric MR data. The Talairach-based brain parcellation method originally proposed by Andreasen and colleagues has several potential advantages over sulcalbased/manual approaches. Most importantly, computer-assisted measurements can be produced quickly, particularly in contrast to time-intensive manual measurements. Second, sulcalbased methods require a high level of neuroanatomical expertise, thus potentially limiting the number of trained personnel available to perform the study and making inter-laboratory replication difficult. Third, manual-based measurements of cerebral topography are prone to rater drift, in contrast to the automated parcellation technique discussed here which can be implemented in a stable manner with high inter-rater and test-retest reliability. Finally, utilization of an atlas-based

D.M. Kaplan et al. / Psychiatry Research: Neuroimaging 76 (1997) 15-27

approach provides a format for describing the relationship between a specified stereotactic coordinate system and neuroanatomic nomenclature, permitting the investigator to quantify, localize and interact with brain space while maintaining neuroanatomic labels. In addition stereotactic coordinate systems provide a framework for multisubject comparisons as well as the potential for correlations between imaging modalities through positional normalization. Although automated atlas-based parcellation methods offer many desirable features, there are several potential circumstances in which sulcalbased techniques may provide methodological advantages. For example, topographic parcellation of cortical regions is likely to provide a more accurate and consistent subdivision of cytoarchitectonic regions and thus more accurate structure-function correspondence. Other drawbacks of the automated method include its difficulty in adjusting for anatomical variation (the Talairach coordinates were fixed on a single elderly female brain), and the assumption of proportionality and left-right symmetry. Finally a problem pertaining to both manual and automated methods is the lack of consensus regarding neuroanatomical boundaries; here, for example, between brainstem and cerebellum. Thus, manual tracing and automated techniques such as the one described here each have their own inherent strengths and weaknesses. As the data set increases in size, the automated technique becomes increasingly advantageous with its speed and consistency. In general, manual tracing continues to have a role for studies requiring more precise sulcal identification or brain structures with complex boundaries. This study describes an automated method sufficiently accurate to detect known anatomical differences in individuals with fragile X and it opens the way for validation of the method with other pediatric neuropsychiatric and neurodevelopmental disorders. Automated methods for quantitative study of pediatric brain structure and function should facilitate the growth of the still limited knowledge base that exists for brain-behavior correlations among normal children as well as children with cognitive, language and behav-

25

ioral disabilities. Current work in our laboratory is focusing on validation of Talairach-based definitions of cerebral lobes and the examination of smaller cortical delineations such as the cingulate cortex, orbitofrontal cortex and dorsolateral prefrontal cortex, all regions that are potentially relevant to understanding disorders of emotion, attention and cognition in children. Acknowledgements

This research was supported by Grants HD31715 (NICHD Human Brain Project), MH00142, NS35359, HD24061 and MH50047. Appendix 1 Revised Talairach sectors for ventricles (lateral, third and velum interpositum)

CaL7, CaL8, DaL6, DaL7, DaL8, ElaL6, ElaL7, ElaL8, E2aL6, E2aL7, E3aL6, E3aL7, FaL6, FaL7, FbL6, FbL7, FbL8, FbL9, GbL7, GbL8, GbL9; CAR7, CAR8, DaR6, DaR7, DaR8, ElaR6, ElaR7, ElaR8, E2aR6, E2aR7, E3aR6, E3aR7, FAR6, FAR7, FbR6, FbR7, FbR8, FbR9, GbR7, GbR8, GbR9. Revised Talairach sectors for cerebellum

IaRll, IaR12, IaR13, IbRll, IbR12, IbR13, IcRll, ICR12, ICR13, HaR10, HaRll, HaR12, HaR13, HaR14, HbRll, HbR12, HbR13, HbR14, HcRll, HcR12, HcR13, HcR14, GaR9, GaR10, GbR10, GaRll, GaR12, GaR13, GaR14, GbR11, GbR12, GbR13, GbR14, GcRll, GcR12, GcR13, GcR14, GdRll, GdR12, GdR13, GdR14, FbRll, FbR12, FbR13, FcRll, FcR12, FcR13IaLll, IaL12, IaL13, IbLll, IbL12, IbL13, IcLll, IcL12, ICL13, HaLlO, HaLll, HAL12, HAL13, HAL14, HbLll, HbL12, HbL13, HbL14, HcLll, HcL12, HcL13, HcL14, GaL9, GaLl0, GbL10, GaLll, GaLl2, GaLl3, GaLl4, GbLll, GbL12, GbL13, GbL14, GcLll, GcL12, GcL13, GcL14, GdLll, GdL12, GdL13, GdL14, FbLll, FbL12, FbL13, FcLll, FcL12, FcL13.

26

D.M. Kaplan et al. / PsychiatryResearch: Neuroimaging 76 (1997) 15-27

Talairach sectors for brainstem

E2aR10, E2aRll, E2aR12, E3aR10, E3aRll, E3aR12, E3aR13, FAR9, FAR10, FAR11, FAR12, FaR13E2aL10, E2aL11, E2aL12, E3aL10, E3aL11, E3aL12, E3aL13, FaL9, FaLl0, FaLl1, FaLl2, FaLl3. Revised Talairach sectors for frontal lobe

AaL10, AaLll, AaL2, AaL3, AaL4, AaLS, AaL6, AaLT, AaL8, AaL9, AaR10, AaRll, AaR2, AaR3, AaR4, AaR5, AaR6, AaR7, AaR8, AaR9, AbL10, AbLll, AbL2, AbL3, AbIA, AbLS,AbL6, AbL7, AbL8, AbL9, AbR10, AbRll, AbR2, AbR3, AbR4, AbRS, AbR6, AbR7, AbR8, AbR9, AcL10, AcLll, AcL2, AcL3, AcL4, AcLS, AcL6, AcL7, AcL8, AcL9, AcR10, AcRll, AcR2, AcR3, AcR4, AcRS, AcR6, AcR7, AcR8, AcR9, AdL10, AdLll, AdL2, AdL3, AdIA, AdLS, AdL6, AdL7, AdL8, AdL9, AdR10, AdRll, AdR2, AdR3, AdR4, AdRS, AdR6, AdR7, AdR8, AdR9, BaLl0, BaLll, BaL2, BaL3, BaIA, BaLS, BaL6, BaL7, BaL8, BaL9, BAR10, BaRll, BAR2, BAR3, BAR4, BAR5, BAR6, BAR7, BAR8, BAR9, BbL10, BbLll, BbL2, BbL3, BbL4, BbLS, BbL6, BbL7, BbL8, BbL9, BbR10, BbRll, BbR2, BbR3, BbR4, BbR5, BbR6, BbR7, BbR8, BbR9, BcL10, BcLll, BcL2, BcL3, BcIA, BcLS, BcL6, BcL7, BcL8, BcL9, BcR10, BcRll, BcR2, BcR3, BcR4, BcR5, BcR6, BcR7, BcR8, BcR9, BdL10, BdLll, BdL2, BdL3, BdLA, BdLS, BdL6, BdL7, BdL8, BdL9, BdR10, BdRll, BdR2, BdR3, BdR4, BdR5, BdR6, BdR7, BdR8, BdR9, CaLl, CaLl0, CaLll, CaL2, CaL3, CaIA, CaLS, CaL6, CaL7, CaL8, CaL9, CAR1, CaRlO, CaRll, CAR2, CAR3, CAR4, CAR5, CAR6, CAR7, CAR8, CAR9, CbL1, CbL10, CbL2, CbL3, CbIA, CbLS, CbL6, CbL7, CbL8, CbL9, CbR1, CbR10, CbR2, CbR3, CbR4, CbRS, CbR6, CbR7, CbR8, CbR9, CcL1, CcL2, CcL3, CcIA, CcL5, CcL6, CcL7, CcL8, CcL9, CcR1, CcR2, CcR3, CcR4, CcR5, CcR6, CcR7, CcR8, CcR9, CdL1, CdL2, CdL3, CdL4, CdLS, CdL6, CdL7, CdL8, CdL9, CdR1, CdR2, CdR3, CdR4, CdRS, CdR6, CdR7, CdR8, CdR9, DaL1, DaL10, DaL2, DaL3, DaIA, DaLS, DaL6, DaR1, DaR10, DaR2, DaR3, DaR4, DaRS, DaR6, DbL1, DbL2, DbL3, DbIA, DbL5, DbL6, DbR1, DbR2, DbR3, DbR4, DbRS,

DbR6, DcL1, DcL2, DcL3, DclA, DcL5, DcL6, DcL7, DcL8, DcR1, DcR2, DcR3, DcR4, DcR5, DcR6, DcR7, DcR8, DdL1, DdL2, DdL3, DdlA, DdL5, DdL6, DdL7, DdL8, DdR1, DdR2, DdR3, DdR4, DdR5, DdR6, DdR7, DdR8, ElaL1, ElaL2, ElaL3, ElalA, ElaL5, ElaR1, ElaR2, ElaR3, ElaR4, ElaR5, ElbL1, ElbL2, ElbL3, ElblA, ElbL5, ElbL6, ElbR1, ElbR2, ElbR3, ElbR4, ElbR5, ElbR6, ElcL1, ElcL2, ElcL3, ElclA, ElcL5, ElcL6, ElcR1, ElcR2, ElcR3, ElcR4, ElcR5, ElcR6, EldL1, EldL2, EldL3, EldL4, EldL5, EldL6, EldL7, EldR1, EldR2, EldR3, EldR4, EldR5, EldR6, EldR7, E2aL1, E2aL2, E2aL3, E2alA, E2aL5, E2aR1, E2aR2, E2aR3, E2aR4, E2aR5, E2bL1, E2bL2, E2bL3, E2bL4, E2bL5, E2bR1, E2bR2, E2bR3, E2bR4, E2bR5, E2cL1, E2cL2, E2cL3, E2clA, E2cR1, E2cR2, E2cR3, E2cR4, E2dL1, E2dL2, E2dL3, E2dR1, E2dR2, E2dR3, E3aL1, E3aL2, E3aL3, E3aL4, E3aL5, E3aL6, E3aR1, E3aR2, E3aR3, E3aR4, E3aR5, E3aR6, E3bL1, E3bL2, E3bL3, E3bL4, E3bR1, E3bR2, E3bR3, E3bR4, E3cL1, E3cL2, E3cR1, E3cR2, FaLl, FaL2, FaL3, FAR1, FAR2, FAR3. References

Andreasen,N.C.,Flashman,L., Flaum,M., Arndt,S., Swayze, V. II, O'Lealy, D.S., Ehrhardt, J.C., Yuh, W.T., 1994a. Regional brain abnormalitiesin schizophreniameasured withmagneticresonanceimaging.Journalof the American MedicalAssociation272, 1763-1769. Andreasen, N.C., Rajaprabhakaran, R., Cizadio, T. (Eds.), 1994b.AutomatedAtlas-BasedDissectionof HumanBrain from MR Images. Society of Magnetic Resonance in Medicine,San Francisco. Andreasen, N.C., Rajarethinam, R., Cizadlo, T., Arndt, S., Swayze,V.W.II, Flashman,L.A.,O'Leary,D.S.,Ehrhardt, J.C., Yuh, W.T., 1996.Automaticatlas-basedvolumeestimation of humanbrain regionsfrom MR images.Journal of ComputorAssistedTomography20, 98-106. Aylward,E.H.,Bran&,J., Codori,A.M.,Mangus,R.S.,Barta, P.E., Harris, G.J., 1994. Reduced basal ganglia volume associatedwiththe genefor Huntington'sdiseasein asymptomatic at riskpersons.Neurology44, 823-828. Aylward,E.H., Reiss,A.L., 1991.Area and volumemeasurement of posterior fossa structures in MRL Journal of PsychiatricResearch25, 159-168. Browne,M.A.,Gaydecki,P.A.,1987.High-speedsplinefitting, with applicationto boundarytracingin low-contrastdigital images.Computorsin Biologyand Medicine17, 109-116.

D.M. Kaplan et al. / Psychiatry Research." Neuroimaging 76 (1997) 15-27 Jernigan, T.L., Bellugi, U., Sowell, E., Doherty, S., Hesselink, J.R., 1993. Cerebral morphologic distinctions between Williams and Down syndromes. Archives of Neurology 50, 186-191. Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man and Cybernetics 9, 62-66. Rademacher, J., Galaburda, A.M., Kennedy, D.N., Filipek, P.A., Caviness, V.S., 1992. Human cerebral cortex: localization, parcellation and morphometry with magnetic resonance imaging. Journal of Cognitive Neuroscience 4, 352-374. Reiss, A.L., Lee, J., Freund, L., 1994. Neuranatomy of fragile X syndrome: the temporal lobe. Neurology 44, 1317-1324. Reiss, A.L., Abrams, M.T., Greenlaw, R., Freund, L., Denckla, M.B., 1995. Neurodevelopmental effects of the FMR-1 full mutation in humans. Nature Medicine 1, 159-167. Reiss, A.L., 1997. Brainlmage, 2.x ed. Stanford University School of Medicine, Stanford, CA. Rousseau, F., Heitz, D., Biancalana, V., Blumenfeld, S., Kretz, C., Boue, J., Tommerup, N., Van Der Hagen, C., DeLozier-Blanchet, C., Croquette, M.F., Gilgenkrantz, S., Jalbcrt, P., Voelckel, M.A., Oberle, I., Mandel, J.L., 1991.

27

Direct diagnosis by DNA analysis of the fragile X syndrome of mental retardation. New England Journal of Medicine 325, 1673-1681. Singer, H.S., Reiss, A.L., Brown, J.E., Aylward, E.H., Shis, B., Chee, E., Harris, E.L., Reader, M.J., Chase, G.A., Bryan, R.N., Denckia, M.B., 1993. Volumetric MRI changes in basal ganglia of children with Tourette's syndrome. Neurology 43, 950-956. Talairach, J., Toumoux, P., 1988. Co-planar Stereotaxic Atlas of the Human Brain. Thieme Medical Publishers, New York. Thorndike, R., Hagen, E., Sattler, J., 1986. Guide for Administering and Scoring the Fourth Edition Stanford-Binet Intelligence Scale. Riverside Publishing, Chicago. Wechsler, D., 1981. Wechsler Adult Intelligence Scale-Revised. The Psychological Corporation, San Antonio, TX. Wechsler, D., 1991. Wechsler Intelligence Scale for ChildrenIII. The Psychological Corporation, San Antonio, TX. Zipursky, R.B., Lim, K.O., Sullivan, E.V., Brown, B.W., Pfefferbaum, A., 1992. Widespread cerebral gray matter volume deficits in schizophrenia. Archives of General Psychiatry 49, 195-205.