Psychiatry Research: Neuroimaging, 40:1-10
I
Elsevier
Is the VBR Still a Useful Measure of Changes in the Cerebral Ventricles? Bryan T. Woods, Andrew Douglass, and Bryan Gescuk Received January 16, 1990; revised version received September 4, 1990; accepted November I0. 1990. Abstract. Standard methods of ventricular measurement with computed tomography (CT) and magnetic resonance imaging (MRI) allow for as much as a 10-mm variation in the level of the "best" axial slice through the lateral ventricles. In a series of 10 patients who received MRI scans, the effect of I-ram variations in slice center level on measured ventricle-brain ratio (VBR) was determined; even variations as small as 1 mm can result in a 10% change in VBR. The amount of within-subject variability in VBR that could result from this factor alone appears to be more than twice as large as that encountered with direct volume measurements, indicating that the latter measure would be at least twice as sensitive to real changes in ventricular size as the VBR.
Key Words. Computed tomography, magnetic resonance imaging, lateral ventricles, ventricle-brain ratio, variability. Shortly after the advent of computed tomography (CT), Synek et al. (1976) proposed a new measurement ofventricular size that would replace the linear measurements in use with pneumoencephalography. This measure, the ventricle-brain ratio (VBR), used the planimetrically measured area of the lateral ventricles on the axial slice where they are largest as the numerator of the ratio, and the area of the whole brain on the same slice as the denominator, The latter measurement was used to correct for systematic variations in ventricular size due to variations in brain size. When it was subsequently reported that the VBRs of a series of patients with enlarged ventricles correlated much more highly with an independently derived volume measurement than did previously described linear measurements (Penn et al., 1978), the value of the VBR seemed established. Subsequently, the majority of CT studies of ventricular size in psychiatric and neurological disorders have used the VBR. If, however, one looks closely at reported VBRs, it is evident that the measure has considerable variability. Typically in control groups the mean of the VBR is about 0.035 to 0.045 and the standard deviation (SD) is about 0.02 to 0.03 (Weinberger et al., 1979; Andreasen et al., 1982), with a coefficient of variation ( S D / m e a n ratio) of approximately 0.67. The question of how much of this variability is due to true
Bryan T. Woods, M.D., is Director of the McLean Hospital Neurology Department and Brain Imaging Center, and Associate Professor of Neurology at Harvard Medical School. At the time of the study, Andrew Douglass and Bryan Gescuk were Research Assistants in the McLean Hospital Brain Imaging Center and Alcohol Research Center, respectively. (Reprint requests to Dr. B.T. Woods, Dept. of Neurology, McLean Hospital, 115 Mill St., Belmont, MA 02178, USA.) 0165-1781/91/$03.50 © 1991 Elsevier Scientific Publishers Ireland Ltd.
between-subject differences in VBR and how much is due to measurement error has never been systematically addressed. A somewhat similar situation appears to exist for the relatively few studies in which the same subjects have had repeated CT scans to look for systematic changes in ventricular size. Vita et al. (1988) analyzed the results of three "negative" studies in schizophrenic patients (Nasrallah et al., 1986; Illowsky et al., 1988; Vita et al., 1988) and gave the following means and SD's of the change ratio (change in VBR, disregarding sign, from scan 1 to scan 2 = (VBR 2 - VBR 2) . (VBR l ) Mean SD Nasrallah et al. 0.38 0.39 Illowsky et al. 0.20 0.12 Vita et al. 0.24 0.22 Although between-subject differences obviously do not contribute to these changes, the question still remains as to how much of the variability is due to error in the method and how much is due to true fluctuations in individual ventricle size over time. The development of magnetic resonance imaging (MRI) has made direct volumetric measurement of the lateral ventricles more feasible than it was with CT. Nevertheless, since such measures are relatively time-consuming and require special image-processing software, while the VBR is quite easily derived on most standard MRI scanners, one would not wish to abandon the older measure unless direct volumetric measurement were clearly superior. In a previous report, it was pointed out that use of an uncorrected area measurement, such as the VBR, to estimate changes in volume has a systematic bias toward minimizing increases and exaggerating decreases and that for mathematical reasons the correlation coefficient is an inappropriate measure of area-volume relationships (Woods and Matthysse, 1989). The current study addresses a further methodological question about the VBR: What portion of the within-subject variability is attributable to measurement error arising from unavoidable head-position variation from study to study, and as a corollary, how does this error compare to the variability encountered with volumetric measurements of the ventricle?
Methods and Results If one considers the laterally symmetrical but irregular shape of the ventricles, the area transected by a single axial tomographic plane through such a volume will vary not only with the level of the plane in the inferior-superior direction but also its rotation with respect to the three major axes. Standard tomographic technique attempts to eliminate lateral rotation or tilt of the head (i.e., variation around two of the axes) by careful patient positioning, and to standardize rotation around the third axis (head inclination) by relating tomographic planes to a standard external landmark (e.g., the canthomeatal line). Slice level is then varied 8-10 mm at a time with CT scanning; thinner slices (3-5 mm) may be used with MRI. Slices are usually contiguous with CT, but not always with MRI. If slices are contiguous, then the
center of the "best" observed slice can be as much as one-half a slice thickness away in either direction from the center of the true optimal slice (i.e., the slice that would yield the largest ventricular area). The question is, how much difference might either variations in slice level such as these or small variations in head angulation (e.g., + 2-3 ° from the canthomeatal line) make in the observed ventricular area and the calculated VBR? To test the effect of slice level variation, a series of l0 normal volunteers had MRI scans in which head position remained constant and a number of sequential series of contiguous (parallel to the plane of the canthomeatal line) axial slices were acquired over the inferior-superior extent of the lateral ventricles in such a manner that each slice was 3-mm thick (in 6 subjects) or 10-mm thick (in 4 subjects). Each series differed from the immediately preceding series by an offset of 1 mm in elevation, so that when slices from the series were combined, the centers of the now overlapping slices varied by only 1-mm steps over the range examined, while the thickness of the slices was either 3 mm or I0 mm. In two further volunteers, the effects of changes in angle of head tilt were assessed by varying the angle of the axial cuts vis-a-vis an internal reference line connecting the anterior and posterior commissures by 1° at a time while the inferior-superior level (center of rotation of the axial plane) was held constant. All slices were 3-mm thick. All scans were done on a G.E. 1.5 Tesla Signa system with Tl-weighted spin-echo sequences (TR = 500; TE ----20). Once images are acquired, the measurement of the area of any structure on a given image has two potential sources of observer error: boundary-determination errors and tracing errors. It was possible to use a modification of a technique described initially for use with CT images (Pfefferbaum et al., 1986) to (all but) eliminate tracing error. The technique is as follows: The same image is displayed in duplicate on a split screen on the image analysis console, windowed with a full range of gray shades on one side and with a 0-1 range on the other side. The level control for the second (black-white) image which sets the pixel intensity above which all pixels are white, and below which all are black, is then varied until the silhouette of the structure of interest best matches the size and shape of the same structure on the gray-scale image. The intensity level giving the best match thus determines the boundary of the structure. Since image intensity may vary from slice to slice, the boundary threshold is found separately for each slice. Once this threshold is determined, all pixels of higher intensity can be nulled by a software program and the remaining structure is then outlined and measured using a console cursor. Since pixels outside the boundary now "don't count," precision tracing of the remaining structure is not required. The brain border is traced separately on the full-range image: both brain area and ventricle area are then calculated by the system software. Once the boundary threshold has been determined, interrater reliability of the method approaches 0.99. Table 1 shows the actual VBR measurements for 10 subjects for the overlapping slices offset by 1 mm. The number of slices varied since slices were only included if visualization showed them to be at the level of the bodies of the lateral ventricles. Fig. 1 shows nine slices of 3-mm thickness, each offset by I mm from the next; slices 1 and 9 were excluded from the actual analysis; slice 2 was included (but the great cerebral vein seen in the posterior midline just above the third ventricle was, of
4 Table 1. Changes in VBR with 1-mm changes in slice center VBR, 10-mm slices VBR, 3-mm slices Pt. #
140
144
259
511
601
695
183
757
793
518
Slice 1
0.015
0.037
0.036
0.043
0.036
0.060
0,079
0.029
0.059
0.060
Slice 2 Slice 3
0.020 0.021
0.053 0.056
0.039 0.045
0.046 0.052
0.045 0.048
0.065 0.069
-0,099
0.033 0,035
0.058 0,058
0,063 0,067 0,069
Slice 4
0.022
0.058
0.044
0.056
0.048
0.073
0.090
0,038
0,061
Slice 5
0.024
0.053
0.041
0.056
0.050
0.067
0.090
0,032
0,059
0.069
Slice 6 Slice 7
0.024 0.022
0.047 0.050
0.033 --
0.058 0.065
0.047 0.050
0.067 0,063
0.083 0.057
0.028 0.031
0.064 0.062
0,069 0.067
Slice 8 Slice 9
---
0.045 --
---
0.064 0.064
0.047 0.046
0.063 0.059
---
0.026 0.021
0.062 0.062
0.066 0.067
Slice 10
--
--
--
0,059
0.047
0.056
--
--
0.050
-
-
Note. VBR = ventricle-brain ra~o.
Fig. 1. Overlapping 3-rnm-thick axial slices with 1-mm separations between slice centers
The lowest slice is top left (33) and the highest is bottom right (41); Table 1 shows the ventricle-brain ratio (VBR) measurements for the 7 slices from 34 to 40. (The great cerebral vein, which can be seen between the left and right pulvinar on the lowest slices, was excluded from the VBR measurements.)
course, not measured). If more than 10 slices met this criterion, the 10 slices with the largest VBRs were used. It can be seen from Table 1 that even a variation in the level of the center of the slice of as little as 1 mm from that giving the largest VBR can at times result in a > I0% change in VBR. Table 2 shows the SDs and coefficients of variations (SD/mean) of the ventricular areas, cerebral areas, and VBRs for each subject. The mean coefficient of variation for brain area is < 0.01, while that for ventricle and VBR is > 0.10 (range 0.08-0.17). For the subjects with 10-mm thick slices, the pattern of variation is essentially the same as for those with 3-mm slices, indicating that the effects of varying slice level are largely independent of the effects of variation in partial volume effects. Table 3 shows the VBR results for two subjects of 1° changes in the angle of the planes of section relative to a fixed central axis. The measurements for more than one slice of 3-mm thickness are given because as the angle changes, the optimal slice may change. Once again, small changes can result in 8-10% VBR changes and the coefficients of variation of the VBRs (0.08, 0.24) are much larger than the coefficients of variation for the brain areas (0.007, 0.013). These coefficients of variation indicate that the variability of ventricular area with position variation is an order of magnitude greater than that of whole brain area, but the numbers are not directly comparable to the change ratio values from repeated studies noted in the introduction. To make a comparison, the assumption was made that the values for each patient shown in Table 1 constitute the range of "best slice" VBRs that could be expected for that patient on repeated studies using a standard CT technique of 10-ram-thick slices. (The acquired center of the 10-mm-thick best slice could be _+5 mm from the true center before another slice would be the best slice.) If one uses a random number table (Winer, 1971) to select pairs of VBRs from each of the 10 studies of Table l, and designates the first value selected as study 1 and the second as study 2, changes can be calculated just as was done in the schizophrenia studies. When this was done on a single trial, the result was a mean absolute difference in VBR of 0.6 units, with an SD of 0.5 units. Table 4a shows these results, as well as the comparable results for the studies analyzed by Vita et al. (1988), re-expressed as means and SDs. Since the VBR difference due to randomized position variation alone is about one-half the within-subject between-study variance reported for schizophrenia studies, it appears that a large portion of the scan-to-scan differences in those (and other VBR) studies could have arisen simply from scan-toscan position variation. It should be emphasized that the estimates of variability derived by this method are conservative since they consider only variation in slice level and neglect the potential effects of study-to-study variation in head rotation around the three major axes. To put variations of this degree in VBR measurement due to position changes in overall context, preexisting data were examined from two other studies that had directly measured ventricular volume changes over time, using 5-ram-thick contiguous coronal spin-echo sequences (TR = 500; TE = 20). The first study group consisted of 10 college student and hospital employee volunteers (median age 21; range 19o39) whose interscan interval (ISI) averaged 3 months (range 1½ to 6 months). The second group (Kroft et al., 1991) consisted of 10 female chronic alcoholic patients who were in treatment and abstinent at times of scanning.
i'l
~ ~o~ 0 '~" 0
~ d d d d d
il ;,
~o °°
~ ~
0
0
0
0
0
~ o d d d d m
~
0
~
~ d d d d d m
C
~ o d d d d m
~ d d d d
N
(D O
0 ~
0
0
0
0
o d o d d d m
•
0
I.
g_l ~,
~ 0 0 0 0 0
I~
e-
.e
~ 0 0 ~ ~ d d d o
~. ~ o
~
0
~ o o d o d
o ~ g ~
0
0
0
0
~ d d d d d
~ 0 ~ ~ 0 0 0 0 0
E
rr
d
II rr r,n
.~..~ . ~
~
Table 3. Changes in VBR with 1° changes in angle of slice plane Angle change Patient 1 Slice -2 -1 0 +1 +2 1 ( - 3 mm) 0.051 2 0.048 3 (+3 mm) 0.041 VBR mean -- 0.049 VBR SD = 0.004 SD/mean
0.050 0.052 0.049
--- 0.08
0.051 0.052 0.044 Brain mean Brain SD
0.050 0.052 0.049 = 144.78 = 1.878
SD/mean
=
0.053 0.053 0.043
0.013
Angle change Patient 2 Slice
-3
1 0.030 2 (+3 mm) 0.018 VBR mean = 0.021 VBRSD
=0.005
SD/mean = 0.24
-2 0.026 0.025
-1
0
0.016 0.015 0.023 0.021 Brain mean = 163.194 BrainSD
=1.132
SD/mean
= 0.007
+1
+2
+3
0.019 0.022
0.014 0.021
0.017 0.021
Note. VBR = ventricle-brain ratio.
The median age was 28 (range 22-61), and the average ISI was 43 days (range 26-67 days). Ventricular volume measurements used different techniques for the two studies. The first method (A. Douglass, unpublished Honors Thesis), used for the normal volunteers, was an experimental technique designed to minimize partial volume effects on ventricle-boundary determination. It involved plotting a curve for the total number of pixels at each intensity value for a region of interest that included both the lateral ventricles and the surrounding gray and white matter, and then having observers blindly select the point of maximum transition of the curve from cerebrospinal fluid values to tissue values. The second technique, which is identical to that described in the current report for the ventricular area measurements, was used for the alcoholic patients. The net change in volume from scan l to scan 2 was minimal and nonsignificant for both studies. Table 4b shows these results in the same mean and SD format as the VBR results. When the VBR and volume results in Table 4a and 4b are compared, several points are evident: (l) Almost all of the correlations from study 1 to study 2 are quite high. (2) Both types of measurement show roughly similar variation between subjects, as indicated by the ratios of SDs to mean scores, suggesting that the between-subject variances in distribution of VBRs and of volumes in the different groups are similar. Finally, by contrast, the within-subject variation for the volume measurements is half or less of that for the area (VBR) measurements, as shown by the ratios of the SDs of the differences to the average SDs of the means (last column).
,,1 ':'=q
|
> "°St io ;I°°°° •
~/
~
i
/i, ,.loooo d
~
©~ ._=
/~-~ /~ol
I"-
9
Conclusion Efforts to measure brain structures precisely using CT or MRI must confront two major sources of variation: true variation, either between individuals, or over time in the same individual, and variation due to error, whether due to instrumentation, technique, or observer. This study has focused primarily on a single source of technique error: variation in patient positioning. In the case of the VBR measurements, it can be seen that differences in position of 2-3 mm of elevation, or 2-3 ° of rotation, which are extremely difficult to eliminate under clinical conditions, can produce surprisingly large changes in the VBR. The effect on scan-rescan results of this degree of position variation alone can be estimated statistically by a mock experiment in which pairs of slices are randomly selected for each patient, and the group difference calculated. When this is done, the ratio of between-subject variation to within-subject variation is similar to that reported for three recent clinical studies using the VBR to look at ventricular change over time, while by way of contrast, in two separate studies that used direct volume measurements to compare ventricular size over time, the ratio was less than half as large. The obvious explanation for the lesser direct volume measurement variability is that it is based on contiguous nonoverlapping slices covering the whole structure, and is thus relatively insensitive to small variations in position orientation; what is missed on one slice will be picked up on another. This insensitivity to orientation is only relative, since it neglects the interaction of partial volume effects with orientation that arises from the asymmetrical shape of both the voxels (e.g., 1 mm X 1 mm X 5 mm) and the ventricles. Thus, one cannot say how much of the remaining variability of the volumetric measurement is also due to partial volume or other measurement error--only that that due to true intrasubject ventricle volume change over time is much less than might have been expected based on the VBR results above. (It is just because the methods of volumetric measurement used do not deal with partial volume effects in a wholly satisfactory manner that neither method is being proposed as a final standard.) The reduced variability with volumetric measurements, which has previously been reported for CT scans on twin pairs by Reveley (1985), is of practical importance. If one wishes to detect changes in either group or individual ventricular volume over time, then the size of the change that can be detected at a given level of confidence varies directly with the size of the measurement variance (Cohen, 1977). Our data indicate that direct volumetric measurements should be able to detect study-to-study volumetric changes in size of lateral ventricles less than half as large as those detectable at the same confidence levels with standard VBR measurements. In another report (Woods and Matthysse, 1989), a simple correction for the mathematical distortion introduced by using an area ratio to estimate changes in volume has been described. In the case of variability due to position changes from scan to scan, one solution that suggests itself is use of thinner slices. This would reduce variability by reducing the error in "best slice" selection, but controls only for variation in slice level, not variation in slice angulation. Since there are a number of disease entities (e.g., Alzheimer's and Huntington's disease, hydrocephalus, and
10 schizophrenia) in which more sensitive detection of changes in ventricular size would provide valuable clinical information, the limitations of the VBR in comparison to the methods of volumetric measurement used in this study suggest that the answer to the question posed in the title may be negative. The ideal method of volumetric measurement may not yet be available, but even currently existing methods appear to be preferable to the VBR.
Acknowledgment. This research was supported in part by grant #AA06252 from the National Institute on Alcohol Abuse and Alcoholism; grants #DA00064, #DA00101, and #DA04059 from the National Institute on Drug Abuse; grant RR05484 from the National Institutes of Health; and a grant from General Electric Medical Systems, Milwaukee, WI. The authors thank Christine Waternaux, Ph.D., for her statistical advice.
References Andreasen, N.C.; Smith, M.R.; Jacoby, C.G.; Dennert, J.W.; and Olsen, S.A. Ventricular enlargement in schizophrenia: Definition and prevalence. American Journal of Psychiatry, 139:292-302, 1982. Cohen, J. Statistical Power Analysis for the Behavioral Sciences. Rev. ed. New York: Academic Press, 1977. Illowsky, B.P.; Juliano, A.M.; Bigelow, L.B.; and Weinberger, D.R. Stability of CT scan findings in schizophrenia: Results of an eight-year follow-up study. Journal of Neurology, Neurosurgery and Psychiatry, 51:209-213, 1988. Kroft, C.L.; Gescuk, B.; Woods, B.T.; Mello, N.K.; Weiss, R.D.; and Mendelson, J.H. Brain ventricular size in female alcoholics: An MRI study. Alcohol, 8:31-34, 1991. Nasrallah, H.A.; Olson, S.C.; McCalley-Whitters, M.; Chapman S.; and Jacoby, D.G. Cerebral ventricular enlargement in schizophrenia. Archives of General Psychiatry, 43:157159, 1986. Penn, R.D.; Belanger, M.G.; and Yasnoff, W.A. Ventricular volume in man computed from CAT scans. Annals of Neurology, 3:216-223, 1978. Pfefferbaum, A.; Zatz, L.M.; and Jernigan, T.H. Computer-interactive method for quantifying cerebrospinal fluid and tissue in brain CT scans: Effects of aging. Journal of Computer Assisted Tomography, 10:571-578, 1986. Reveley, M.A. Ventricular enlargement in schizophrenia: The validity of computerized tomographic findings. British Journal of Psychiatry, 147:233-240, 1985. Synek, V.; Reuben, J.R.; and Du Boulay, G.H. Comparing Evan's index and computerized axial tomography in assessing relationship of ventricle size to brain size. Neurology, 26:231233, 1976. Vita, A.; Sacchetti, E.; Valvasori, G.; and Cazzullo, C.L. Brain morphology in schizophrenia: A 2- to 5-year CT scan follow-up study. Acta Psychiatrica Scandinavica, 78:618-621, 1988. Weinberger, D.R.; Torrey, E.F.; Neophytides, A.N.; and Wyatt, R.J. Lateral cerebral ventricular enlargement in chronic schizophrenia. Archives of General Psychiatry, 36:735-739, 1979. Winer, B.J. Statistical Principles in Experimental Design, 2nd ed. New York: McGraw-Hill, 1971. Woods, B.T., and Matthysse, S. A simple method for conversion of area ratio measurements of cerebral structures to volume estimates. Biological Psychiatry, 26:748-752, 1989.
I
Elsevier
Is the VBR Still a Useful Measure of Changes in the Cerebral Ventricles? Bryan T. Woods, Andrew Douglass, and Bryan Gescuk Received January 16, 1990; revised version received September 4, 1990; accepted November I0. 1990. Abstract. Standard methods of ventricular measurement with computed tomography (CT) and magnetic resonance imaging (MRI) allow for as much as a 10-mm variation in the level of the "best" axial slice through the lateral ventricles. In a series of 10 patients who received MRI scans, the effect of I-ram variations in slice center level on measured ventricle-brain ratio (VBR) was determined; even variations as small as 1 mm can result in a 10% change in VBR. The amount of within-subject variability in VBR that could result from this factor alone appears to be more than twice as large as that encountered with direct volume measurements, indicating that the latter measure would be at least twice as sensitive to real changes in ventricular size as the VBR.
Key Words. Computed tomography, magnetic resonance imaging, lateral ventricles, ventricle-brain ratio, variability. Shortly after the advent of computed tomography (CT), Synek et al. (1976) proposed a new measurement ofventricular size that would replace the linear measurements in use with pneumoencephalography. This measure, the ventricle-brain ratio (VBR), used the planimetrically measured area of the lateral ventricles on the axial slice where they are largest as the numerator of the ratio, and the area of the whole brain on the same slice as the denominator, The latter measurement was used to correct for systematic variations in ventricular size due to variations in brain size. When it was subsequently reported that the VBRs of a series of patients with enlarged ventricles correlated much more highly with an independently derived volume measurement than did previously described linear measurements (Penn et al., 1978), the value of the VBR seemed established. Subsequently, the majority of CT studies of ventricular size in psychiatric and neurological disorders have used the VBR. If, however, one looks closely at reported VBRs, it is evident that the measure has considerable variability. Typically in control groups the mean of the VBR is about 0.035 to 0.045 and the standard deviation (SD) is about 0.02 to 0.03 (Weinberger et al., 1979; Andreasen et al., 1982), with a coefficient of variation ( S D / m e a n ratio) of approximately 0.67. The question of how much of this variability is due to true
Bryan T. Woods, M.D., is Director of the McLean Hospital Neurology Department and Brain Imaging Center, and Associate Professor of Neurology at Harvard Medical School. At the time of the study, Andrew Douglass and Bryan Gescuk were Research Assistants in the McLean Hospital Brain Imaging Center and Alcohol Research Center, respectively. (Reprint requests to Dr. B.T. Woods, Dept. of Neurology, McLean Hospital, 115 Mill St., Belmont, MA 02178, USA.) 0165-1781/91/$03.50 © 1991 Elsevier Scientific Publishers Ireland Ltd.
between-subject differences in VBR and how much is due to measurement error has never been systematically addressed. A somewhat similar situation appears to exist for the relatively few studies in which the same subjects have had repeated CT scans to look for systematic changes in ventricular size. Vita et al. (1988) analyzed the results of three "negative" studies in schizophrenic patients (Nasrallah et al., 1986; Illowsky et al., 1988; Vita et al., 1988) and gave the following means and SD's of the change ratio (change in VBR, disregarding sign, from scan 1 to scan 2 = (VBR 2 - VBR 2) . (VBR l ) Mean SD Nasrallah et al. 0.38 0.39 Illowsky et al. 0.20 0.12 Vita et al. 0.24 0.22 Although between-subject differences obviously do not contribute to these changes, the question still remains as to how much of the variability is due to error in the method and how much is due to true fluctuations in individual ventricle size over time. The development of magnetic resonance imaging (MRI) has made direct volumetric measurement of the lateral ventricles more feasible than it was with CT. Nevertheless, since such measures are relatively time-consuming and require special image-processing software, while the VBR is quite easily derived on most standard MRI scanners, one would not wish to abandon the older measure unless direct volumetric measurement were clearly superior. In a previous report, it was pointed out that use of an uncorrected area measurement, such as the VBR, to estimate changes in volume has a systematic bias toward minimizing increases and exaggerating decreases and that for mathematical reasons the correlation coefficient is an inappropriate measure of area-volume relationships (Woods and Matthysse, 1989). The current study addresses a further methodological question about the VBR: What portion of the within-subject variability is attributable to measurement error arising from unavoidable head-position variation from study to study, and as a corollary, how does this error compare to the variability encountered with volumetric measurements of the ventricle?
Methods and Results If one considers the laterally symmetrical but irregular shape of the ventricles, the area transected by a single axial tomographic plane through such a volume will vary not only with the level of the plane in the inferior-superior direction but also its rotation with respect to the three major axes. Standard tomographic technique attempts to eliminate lateral rotation or tilt of the head (i.e., variation around two of the axes) by careful patient positioning, and to standardize rotation around the third axis (head inclination) by relating tomographic planes to a standard external landmark (e.g., the canthomeatal line). Slice level is then varied 8-10 mm at a time with CT scanning; thinner slices (3-5 mm) may be used with MRI. Slices are usually contiguous with CT, but not always with MRI. If slices are contiguous, then the
center of the "best" observed slice can be as much as one-half a slice thickness away in either direction from the center of the true optimal slice (i.e., the slice that would yield the largest ventricular area). The question is, how much difference might either variations in slice level such as these or small variations in head angulation (e.g., + 2-3 ° from the canthomeatal line) make in the observed ventricular area and the calculated VBR? To test the effect of slice level variation, a series of l0 normal volunteers had MRI scans in which head position remained constant and a number of sequential series of contiguous (parallel to the plane of the canthomeatal line) axial slices were acquired over the inferior-superior extent of the lateral ventricles in such a manner that each slice was 3-mm thick (in 6 subjects) or 10-mm thick (in 4 subjects). Each series differed from the immediately preceding series by an offset of 1 mm in elevation, so that when slices from the series were combined, the centers of the now overlapping slices varied by only 1-mm steps over the range examined, while the thickness of the slices was either 3 mm or I0 mm. In two further volunteers, the effects of changes in angle of head tilt were assessed by varying the angle of the axial cuts vis-a-vis an internal reference line connecting the anterior and posterior commissures by 1° at a time while the inferior-superior level (center of rotation of the axial plane) was held constant. All slices were 3-mm thick. All scans were done on a G.E. 1.5 Tesla Signa system with Tl-weighted spin-echo sequences (TR = 500; TE ----20). Once images are acquired, the measurement of the area of any structure on a given image has two potential sources of observer error: boundary-determination errors and tracing errors. It was possible to use a modification of a technique described initially for use with CT images (Pfefferbaum et al., 1986) to (all but) eliminate tracing error. The technique is as follows: The same image is displayed in duplicate on a split screen on the image analysis console, windowed with a full range of gray shades on one side and with a 0-1 range on the other side. The level control for the second (black-white) image which sets the pixel intensity above which all pixels are white, and below which all are black, is then varied until the silhouette of the structure of interest best matches the size and shape of the same structure on the gray-scale image. The intensity level giving the best match thus determines the boundary of the structure. Since image intensity may vary from slice to slice, the boundary threshold is found separately for each slice. Once this threshold is determined, all pixels of higher intensity can be nulled by a software program and the remaining structure is then outlined and measured using a console cursor. Since pixels outside the boundary now "don't count," precision tracing of the remaining structure is not required. The brain border is traced separately on the full-range image: both brain area and ventricle area are then calculated by the system software. Once the boundary threshold has been determined, interrater reliability of the method approaches 0.99. Table 1 shows the actual VBR measurements for 10 subjects for the overlapping slices offset by 1 mm. The number of slices varied since slices were only included if visualization showed them to be at the level of the bodies of the lateral ventricles. Fig. 1 shows nine slices of 3-mm thickness, each offset by I mm from the next; slices 1 and 9 were excluded from the actual analysis; slice 2 was included (but the great cerebral vein seen in the posterior midline just above the third ventricle was, of
4 Table 1. Changes in VBR with 1-mm changes in slice center VBR, 10-mm slices VBR, 3-mm slices Pt. #
140
144
259
511
601
695
183
757
793
518
Slice 1
0.015
0.037
0.036
0.043
0.036
0.060
0,079
0.029
0.059
0.060
Slice 2 Slice 3
0.020 0.021
0.053 0.056
0.039 0.045
0.046 0.052
0.045 0.048
0.065 0.069
-0,099
0.033 0,035
0.058 0,058
0,063 0,067 0,069
Slice 4
0.022
0.058
0.044
0.056
0.048
0.073
0.090
0,038
0,061
Slice 5
0.024
0.053
0.041
0.056
0.050
0.067
0.090
0,032
0,059
0.069
Slice 6 Slice 7
0.024 0.022
0.047 0.050
0.033 --
0.058 0.065
0.047 0.050
0.067 0,063
0.083 0.057
0.028 0.031
0.064 0.062
0,069 0.067
Slice 8 Slice 9
---
0.045 --
---
0.064 0.064
0.047 0.046
0.063 0.059
---
0.026 0.021
0.062 0.062
0.066 0.067
Slice 10
--
--
--
0,059
0.047
0.056
--
--
0.050
-
-
Note. VBR = ventricle-brain ra~o.
Fig. 1. Overlapping 3-rnm-thick axial slices with 1-mm separations between slice centers
The lowest slice is top left (33) and the highest is bottom right (41); Table 1 shows the ventricle-brain ratio (VBR) measurements for the 7 slices from 34 to 40. (The great cerebral vein, which can be seen between the left and right pulvinar on the lowest slices, was excluded from the VBR measurements.)
course, not measured). If more than 10 slices met this criterion, the 10 slices with the largest VBRs were used. It can be seen from Table 1 that even a variation in the level of the center of the slice of as little as 1 mm from that giving the largest VBR can at times result in a > I0% change in VBR. Table 2 shows the SDs and coefficients of variations (SD/mean) of the ventricular areas, cerebral areas, and VBRs for each subject. The mean coefficient of variation for brain area is < 0.01, while that for ventricle and VBR is > 0.10 (range 0.08-0.17). For the subjects with 10-mm thick slices, the pattern of variation is essentially the same as for those with 3-mm slices, indicating that the effects of varying slice level are largely independent of the effects of variation in partial volume effects. Table 3 shows the VBR results for two subjects of 1° changes in the angle of the planes of section relative to a fixed central axis. The measurements for more than one slice of 3-mm thickness are given because as the angle changes, the optimal slice may change. Once again, small changes can result in 8-10% VBR changes and the coefficients of variation of the VBRs (0.08, 0.24) are much larger than the coefficients of variation for the brain areas (0.007, 0.013). These coefficients of variation indicate that the variability of ventricular area with position variation is an order of magnitude greater than that of whole brain area, but the numbers are not directly comparable to the change ratio values from repeated studies noted in the introduction. To make a comparison, the assumption was made that the values for each patient shown in Table 1 constitute the range of "best slice" VBRs that could be expected for that patient on repeated studies using a standard CT technique of 10-ram-thick slices. (The acquired center of the 10-mm-thick best slice could be _+5 mm from the true center before another slice would be the best slice.) If one uses a random number table (Winer, 1971) to select pairs of VBRs from each of the 10 studies of Table l, and designates the first value selected as study 1 and the second as study 2, changes can be calculated just as was done in the schizophrenia studies. When this was done on a single trial, the result was a mean absolute difference in VBR of 0.6 units, with an SD of 0.5 units. Table 4a shows these results, as well as the comparable results for the studies analyzed by Vita et al. (1988), re-expressed as means and SDs. Since the VBR difference due to randomized position variation alone is about one-half the within-subject between-study variance reported for schizophrenia studies, it appears that a large portion of the scan-to-scan differences in those (and other VBR) studies could have arisen simply from scan-toscan position variation. It should be emphasized that the estimates of variability derived by this method are conservative since they consider only variation in slice level and neglect the potential effects of study-to-study variation in head rotation around the three major axes. To put variations of this degree in VBR measurement due to position changes in overall context, preexisting data were examined from two other studies that had directly measured ventricular volume changes over time, using 5-ram-thick contiguous coronal spin-echo sequences (TR = 500; TE = 20). The first study group consisted of 10 college student and hospital employee volunteers (median age 21; range 19o39) whose interscan interval (ISI) averaged 3 months (range 1½ to 6 months). The second group (Kroft et al., 1991) consisted of 10 female chronic alcoholic patients who were in treatment and abstinent at times of scanning.
i'l
~ ~o~ 0 '~" 0
~ d d d d d
il ;,
~o °°
~ ~
0
0
0
0
0
~ o d d d d m
~
0
~
~ d d d d d m
C
~ o d d d d m
~ d d d d
N
(D O
0 ~
0
0
0
0
o d o d d d m
•
0
I.
g_l ~,
~ 0 0 0 0 0
I~
e-
.e
~ 0 0 ~ ~ d d d o
~. ~ o
~
0
~ o o d o d
o ~ g ~
0
0
0
0
~ d d d d d
~ 0 ~ ~ 0 0 0 0 0
E
rr
d
II rr r,n
.~..~ . ~
~
Table 3. Changes in VBR with 1° changes in angle of slice plane Angle change Patient 1 Slice -2 -1 0 +1 +2 1 ( - 3 mm) 0.051 2 0.048 3 (+3 mm) 0.041 VBR mean -- 0.049 VBR SD = 0.004 SD/mean
0.050 0.052 0.049
--- 0.08
0.051 0.052 0.044 Brain mean Brain SD
0.050 0.052 0.049 = 144.78 = 1.878
SD/mean
=
0.053 0.053 0.043
0.013
Angle change Patient 2 Slice
-3
1 0.030 2 (+3 mm) 0.018 VBR mean = 0.021 VBRSD
=0.005
SD/mean = 0.24
-2 0.026 0.025
-1
0
0.016 0.015 0.023 0.021 Brain mean = 163.194 BrainSD
=1.132
SD/mean
= 0.007
+1
+2
+3
0.019 0.022
0.014 0.021
0.017 0.021
Note. VBR = ventricle-brain ratio.
The median age was 28 (range 22-61), and the average ISI was 43 days (range 26-67 days). Ventricular volume measurements used different techniques for the two studies. The first method (A. Douglass, unpublished Honors Thesis), used for the normal volunteers, was an experimental technique designed to minimize partial volume effects on ventricle-boundary determination. It involved plotting a curve for the total number of pixels at each intensity value for a region of interest that included both the lateral ventricles and the surrounding gray and white matter, and then having observers blindly select the point of maximum transition of the curve from cerebrospinal fluid values to tissue values. The second technique, which is identical to that described in the current report for the ventricular area measurements, was used for the alcoholic patients. The net change in volume from scan l to scan 2 was minimal and nonsignificant for both studies. Table 4b shows these results in the same mean and SD format as the VBR results. When the VBR and volume results in Table 4a and 4b are compared, several points are evident: (l) Almost all of the correlations from study 1 to study 2 are quite high. (2) Both types of measurement show roughly similar variation between subjects, as indicated by the ratios of SDs to mean scores, suggesting that the between-subject variances in distribution of VBRs and of volumes in the different groups are similar. Finally, by contrast, the within-subject variation for the volume measurements is half or less of that for the area (VBR) measurements, as shown by the ratios of the SDs of the differences to the average SDs of the means (last column).
,,1 ':'=q
|
> "°St io ;I°°°° •
~/
~
i
/i, ,.loooo d
~
©~ ._=
/~-~ /~ol
I"-
9
Conclusion Efforts to measure brain structures precisely using CT or MRI must confront two major sources of variation: true variation, either between individuals, or over time in the same individual, and variation due to error, whether due to instrumentation, technique, or observer. This study has focused primarily on a single source of technique error: variation in patient positioning. In the case of the VBR measurements, it can be seen that differences in position of 2-3 mm of elevation, or 2-3 ° of rotation, which are extremely difficult to eliminate under clinical conditions, can produce surprisingly large changes in the VBR. The effect on scan-rescan results of this degree of position variation alone can be estimated statistically by a mock experiment in which pairs of slices are randomly selected for each patient, and the group difference calculated. When this is done, the ratio of between-subject variation to within-subject variation is similar to that reported for three recent clinical studies using the VBR to look at ventricular change over time, while by way of contrast, in two separate studies that used direct volume measurements to compare ventricular size over time, the ratio was less than half as large. The obvious explanation for the lesser direct volume measurement variability is that it is based on contiguous nonoverlapping slices covering the whole structure, and is thus relatively insensitive to small variations in position orientation; what is missed on one slice will be picked up on another. This insensitivity to orientation is only relative, since it neglects the interaction of partial volume effects with orientation that arises from the asymmetrical shape of both the voxels (e.g., 1 mm X 1 mm X 5 mm) and the ventricles. Thus, one cannot say how much of the remaining variability of the volumetric measurement is also due to partial volume or other measurement error--only that that due to true intrasubject ventricle volume change over time is much less than might have been expected based on the VBR results above. (It is just because the methods of volumetric measurement used do not deal with partial volume effects in a wholly satisfactory manner that neither method is being proposed as a final standard.) The reduced variability with volumetric measurements, which has previously been reported for CT scans on twin pairs by Reveley (1985), is of practical importance. If one wishes to detect changes in either group or individual ventricular volume over time, then the size of the change that can be detected at a given level of confidence varies directly with the size of the measurement variance (Cohen, 1977). Our data indicate that direct volumetric measurements should be able to detect study-to-study volumetric changes in size of lateral ventricles less than half as large as those detectable at the same confidence levels with standard VBR measurements. In another report (Woods and Matthysse, 1989), a simple correction for the mathematical distortion introduced by using an area ratio to estimate changes in volume has been described. In the case of variability due to position changes from scan to scan, one solution that suggests itself is use of thinner slices. This would reduce variability by reducing the error in "best slice" selection, but controls only for variation in slice level, not variation in slice angulation. Since there are a number of disease entities (e.g., Alzheimer's and Huntington's disease, hydrocephalus, and
10 schizophrenia) in which more sensitive detection of changes in ventricular size would provide valuable clinical information, the limitations of the VBR in comparison to the methods of volumetric measurement used in this study suggest that the answer to the question posed in the title may be negative. The ideal method of volumetric measurement may not yet be available, but even currently existing methods appear to be preferable to the VBR.
Acknowledgment. This research was supported in part by grant #AA06252 from the National Institute on Alcohol Abuse and Alcoholism; grants #DA00064, #DA00101, and #DA04059 from the National Institute on Drug Abuse; grant RR05484 from the National Institutes of Health; and a grant from General Electric Medical Systems, Milwaukee, WI. The authors thank Christine Waternaux, Ph.D., for her statistical advice.
References Andreasen, N.C.; Smith, M.R.; Jacoby, C.G.; Dennert, J.W.; and Olsen, S.A. Ventricular enlargement in schizophrenia: Definition and prevalence. American Journal of Psychiatry, 139:292-302, 1982. Cohen, J. Statistical Power Analysis for the Behavioral Sciences. Rev. ed. New York: Academic Press, 1977. Illowsky, B.P.; Juliano, A.M.; Bigelow, L.B.; and Weinberger, D.R. Stability of CT scan findings in schizophrenia: Results of an eight-year follow-up study. Journal of Neurology, Neurosurgery and Psychiatry, 51:209-213, 1988. Kroft, C.L.; Gescuk, B.; Woods, B.T.; Mello, N.K.; Weiss, R.D.; and Mendelson, J.H. Brain ventricular size in female alcoholics: An MRI study. Alcohol, 8:31-34, 1991. Nasrallah, H.A.; Olson, S.C.; McCalley-Whitters, M.; Chapman S.; and Jacoby, D.G. Cerebral ventricular enlargement in schizophrenia. Archives of General Psychiatry, 43:157159, 1986. Penn, R.D.; Belanger, M.G.; and Yasnoff, W.A. Ventricular volume in man computed from CAT scans. Annals of Neurology, 3:216-223, 1978. Pfefferbaum, A.; Zatz, L.M.; and Jernigan, T.H. Computer-interactive method for quantifying cerebrospinal fluid and tissue in brain CT scans: Effects of aging. Journal of Computer Assisted Tomography, 10:571-578, 1986. Reveley, M.A. Ventricular enlargement in schizophrenia: The validity of computerized tomographic findings. British Journal of Psychiatry, 147:233-240, 1985. Synek, V.; Reuben, J.R.; and Du Boulay, G.H. Comparing Evan's index and computerized axial tomography in assessing relationship of ventricle size to brain size. Neurology, 26:231233, 1976. Vita, A.; Sacchetti, E.; Valvasori, G.; and Cazzullo, C.L. Brain morphology in schizophrenia: A 2- to 5-year CT scan follow-up study. Acta Psychiatrica Scandinavica, 78:618-621, 1988. Weinberger, D.R.; Torrey, E.F.; Neophytides, A.N.; and Wyatt, R.J. Lateral cerebral ventricular enlargement in chronic schizophrenia. Archives of General Psychiatry, 36:735-739, 1979. Winer, B.J. Statistical Principles in Experimental Design, 2nd ed. New York: McGraw-Hill, 1971. Woods, B.T., and Matthysse, S. A simple method for conversion of area ratio measurements of cerebral structures to volume estimates. Biological Psychiatry, 26:748-752, 1989.
