Journal of Cerebral Blood Flow and Metabolism
11:A83-A88
©
1991 The International Society of Cerebral Blood Flow and Metabolism
Effects of Percent Thresholding on the Extraction of
eSP]Pluorodeoxyglucose Positron Emission Tomographic Region-of-Interest Data
D. A. Rottenberg, J. R. Moeller, S. C. Strother, V. Dhawan, and M. L. Sergi Department of Neurology, Memorial Sloan-Kettering Cancer Center, New York, New York, U.S.A.
Summary: Although we and others have employed a
determined by a number of factors, including the relative magnitude of regional activation, ROI size, and the spe cific threshold selected. The difference-annulus concept is proposed as a means to study the effects of different region drawing and thresholding strategies, and to deter mine if a given ROI contains one and only one source of covarying metabolic activity. Key Words: Positron emis sion tomography (PET)-Thresholding-Region of in terest (ROI)-Regional cerebral glucose metabolism (rCMRg1J·
thresholding strategy to extract "peak" values from pos itron emission tomographic (PET) regions of interest (ROIs), the effects of peak picking on fitted fluorodeox yglucose rate constants, regional metabolic rate for glu cose (rCMRglc) profiles, patterns of regional metabolic covariation, and PET-neurobehavioral correlations have not been systematically investigated. Our results suggest that under some commonly encountered imaging condi tions percent thresholding may increase sensitivity to re gional activation; however, the effect of thresholding is
We and others have employed a thresholding strategy to extract stable "peak" values from pos itron emission tomographic (PET) regions of inter est (ROls); however, no one, to our knowledge, has systematically investigated the effects of peak picking on fitted fluorodeoxyglucose (FDG) rate constants, regional metabolic rate for glucose (rCMRglc) profiles, patterns of regional metabolic covariation, or PET-neurobehavioral correlations. Strother et al. (1987) previously concluded that thresholding did not affect the distribution of co variance-pattern weights in a small group of normal volunteer subjects.
Address correspondence and reprint requests to Dr. D. A. Rottenberg at PET Imaging Service (lIP), Veterans Administra tion Medical Center, Minneapolis, MN 55417, U. S.A. Abbreviations used: COV, coefficient of variation; FDG, flu orodeoxyglucose; FWHM, full width at half-maximum; GIS, group invariant subprofile; GM, gray matter; PET, positron emission tomography; rCBV, regional cerebral blood volume; ROI, region of interest; SSF, subprofile scaling factor; SSM, scaled subprofile model.
FIG. 1. The effect of variously thresholding raw-count FDG/PET images. Percent thresholds are indicated below and to the left of each image. Note that RDls are drawn with out regard to exact anatomical boundaries and that thresh olded voxels are not always contiguous.
A83
D. A. ROTTENBERG ET AL.
A84
WHAT IS THRESHOLDING?
Thresholding refers to the selection of voxels with numerical values that fall within a predeter mined range, specified by lower and upper bound ary values, e.g., [Vmax
'
(100 - 1)/100]
�
V
�
Vmax
where Vmax refers to the maximum voxel value, v to a thresholded voxel value, and 1 to the upper per cent threshold. Within the context of this discussion, a 5% threshold value for a given ROI signifies that the A
2.88 1.75 ::Jl > U
.... ....
+------t 5
mm
WHY THRESHOLD?
Because of the limited spatial resolution of PET cameras [0.5-1.0 mm full width at half-maximum (FWHM)] relative to the width of the human corti-
FWHM
1.5. +------+ 5
1.50
: , ,
1.25
'
,
B.75
U
U
....
10
mm
Distance
FWHM
1.5e +------t 10
--I II II II I I I I
1.5.
1.25
mm
FWHM
1.25
>
1.••
U
1:5
(l) 0:::
0.25
+------t
.... ....
13.50
8.8e +--"'----'--"---+
Di stcmce
1.75
13.75
(l) >
13.50
2 •••
1.130
-----1
+__"" 8.88 +-_L----'-_�_�_'_____'__________
B
FWHM
::Jl >
1.ee
mm
1.25
---J
1:5
(l) 0:::
upper 5% ofROI voxel values are averaged and that the average value is assigned to the ROI. Figure 1 illustrates the effect of various percent thresholds on raw-count eSF]fluorodeoxyglucose (FDG)/PET images. Note that thresholding permits a relaxation of stringent spatial constraints: ROls can be drawn quickly, without regard to exact anatomical bound aries, and, within a given ROI, thresholded voxels need not be contiguous.
... ....
•• 5.
1:5
(l) 0:::
•.25
•.••
- ---'_---'-L_--'-_-'-_-'-__-=____+ -/""'-_ Distance
FIG.2. The effect of spatial resolution on peak and mean ROI values. Idealized slabs of gray matter in a cortical convolu tion have been imaged with 5 mm (A) or 10 mm (B) [full width at half-maximum (FWHM)] resolution. The underlying corti cal structure (4 mm cortex, 1 mm gap, 4 mm cortex, 5 mm gap, 5 mm cortex with superimposed 2 mm focus of 50% activation) is indicated by dotted lines and the reconstructed activity profile by solid lines.
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
9.50
8.25
e.ee -F------=�--1 Distance FIG. 3. The effect of increasing levels of cortical activati()n c
5 mm (A) and 10 mm (B) FWHM resolution. In each case, th five curves represent (from lowest to highest) 0, 25, 50, 7! and 100% activation. In B, note that the activated right pea lies below the unactivated left peak until the level of actiVe tion reaches 100%.
AS5
PERCENT THRESHOLDING cal ribbon (�3 mm), discrete foci of functional ac tivation are blurred in a more or less Gaussian fash ion. In some situations, thresholding increases sen sitivity to regional activation effects (Figs. 2-4); however, the "blurring" that results from volume averaging with adjacent white matter and/or nonac tivated cortex cannot be removed or minimized by thresholding techniques. The potential benefits to be derived from thresholding depend ultimately on the macroscopic structure of the gray matter (GM) 1.1 +-----+----+---+
A
5
mm
B.9
....
U
mm
spatial resolution
FWHM
loB
....
TABLE 1. Percent thresholding with 5
B.8
(l) >
METHODS AND RESULTS
.... +'
0
B.7
(l) 0::: B.6
B.5 25
5B
75
lBB
Per-cent Ac t i vcd ion B
Effect of thresholding on fitted FDG rate constants
In order to investigate the effect of thresholding on fitted FDG rate constants, 21 serial PET images (10 x 1 min, 5 x 2 min, 3 x 5 min, 3 x 10 min) were acquired with the PC4600 Positron Camera (Kearfott and Carroll, 1984) following the injection of 5-10 mCi of FDG, and the time course of plasma 18F radioactivity was determined by sampling radial arterial blood. Twenty-eight standardized cortical and subcortical ROIs were outlined on recon structed PET brain slices with reference to a neuroana tomical atlas (E ycleshymer and Schoemaker, 1911; Anderson et a!., 1988), and regional GM rate constants (k1-k3) and cerebral blood volume (rCBV) were estimated from the time course of blood and regional brain radio activity (Evans et a!. , 1986). Ten, 15, and 25% thresholds were used to "fix" ROI voxe1s on 45-to-55-min PET
B.8
10
mm
FWHM
B.7 ::JI ....
>
B.7
....
U ....
0
FDG, produced by a modification of Tewson's synthe sis (Tewson, 1983; Ginos et aI. , 1987), was more than 97% radiochemically pure (specific activity of 500 mCilmmol). Twelve demented AIDS patients and 18 normal volunteer subjects were scanned in a standard resting state with eyes patched and with minimal auditory stimulation (Rot tenberg et aI. , 1987; Anderson et aI. , 1988).
B.6
......
(l) 0::: B.5
TABLE 2. Percent thresholding with 10
mm
spatial resolution
B.5 25
5B
75
lB0
Percent activation
Per-cent Ac t 1 va t ion FIG. 4. The effect of percent thresholding at various levels of cortical activation. The left (A) and right (8) panels are based on the data in Tables 1 and 2, respectively. In each case, the dotted line represents the mean regional value (obtained by averaging all reconstructed voxel values within the outlines of the underlying cortical structure), and the solid lines rep resent the variously thresholded ROI values; the ordering of the percent-threshold profiles (i.e., from 5 to 25%) is indi cated in both panels. Note that thresholding increases our ability to detect activation effects at 5 mm resolution but not at 10 mm resolution.
ROI mean 5% THR 10% THR 15% THR 20% THR 25% THR
0%
25%
50%
75%
100%
0.527 0.620 0.611 0.596 0.579 0.541
0.549 0.623 0.612 0.586 0.564 0.557
0.570 0.625 0.602 0.586 0.582 0.577
0.592 0.627 0.612 0.609 0.602 0.596
0.614 0.666 0.643 0.635 0.631 0.622
All values represent relative activity and refer to plots in Fig. 4B. THR: threshold.
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
A86
D. A. ROTTENBERG ET AL. TABLE 3. Effect of percent thresholding on fitted values of krk3' rCBV, and rCMRgic Percent threshold 10% 0.084 0.108 0.082 0.079
k J k2 k3 rCBV
8.58
rCMRglc
±
0.017 0.053 0.031 0.026
(20%) (49%) (38%) (32%)
0.083 0.107 0.083 0.077
±
1.57 (18%)
8.92
± ± ±
25%
15%
±
0.015 0.051 0.028 0.026
(18%) (47%) (34%) (34%)
0.077 0.097 0.078 0.076
±
1.20 (13%)
8.18
± ± ±
±
0.015 0.048 0.030 0.025
±
1.47 (18%)
± ± ±
(19%) (49%) (39%) (33%)
n 18 normal volunteer subjects, age 27 ± 5 years (mean ± SO). Units: k -k3' min-J; rCBV, ml/g; rCMRglc, mg min-J 100 ml-J. J All values are means across subjects ± SO (COV) of the within-subject means across all cortical and subcortical gray-matter regions of interest. rCMRglc Cp/LC . k k3/(k2 + k3)' J =
=
images (frame 21), and these voxels were then "re played" on frames 1-20 in order to define the time course of ROI radioactivity. The results of fitting rate constants and rCBV to thresholded ROI data are illustrated in Table 3. Although there is a trend toward lower coefficients of variation (COVs) with a 15% threshold, these data sug gest that percent thresholding has little effect on the pre cision with which mean GM FDG rate constants can be estimated. Effect of thresholding on mean
rCMRgIc profiles
ROI values and
Effect of thresholding on patterns of regional metabolic covariation
With FDG/PET (and 0-15/PET) images, increasing the percent threshold from 5 to 25% increases the number of nonactivated and/or non-gray-matter voxels being aver aged and decreases the mean ROI value in a systematic fashion. To illustrate this phenomenon, 28 standardized 19.9 9.8 0'
8.8
(S) (S)
7.8
...... C
6.8
E ...... 0" E
ROIs were outlined on reconstructed PET brain slices from the normal volunteer subjects, "functionalized" (i.e., converted from cps/voxel to rCMRglc units) accord ing to Phelps et al. (1979); mean rCMRglc values were then calculated across subjects for 5, 10, 15, 20, and 25% thresholds and displayed in the form of metabolic pro files. As is evident from Fig. 5, changing the percent threshold appears to scale the mean rCMRglc profile with out altering its shape.
Next, we investigated the stability of group invariant subprofiles (GISs), patterns of metabolic covariation ex tracted using the scaled subprofile model (SSM, Moeller et aI., 1987). GISs and their respective subject weights (subprofile scaling factors, SSFs) were derived from FDG/PET data sets containing ROI values for 12 patients with AIDS dementia and 17 normal volunteer subjects (Rottenberg et aI., 1987). Both raw-count (cps/voxel) and functionalized (glucose metabolic rate) data sets were an alyzed. Raw-count ROI data were thresholded at 10 and 25%, and functionalized ROI data were thresholded at 5, 10, 15, 20, and 25%; additionally, a 25-20% "difference annulus" was operationally defined as the set of voxels lying outside the 20% threshold but within the 25% threshold (Fig. 6). Using multiple linear regression, sub-
�.8
/�'�,''.
•. 8 ::J ...... CJ 0:: L U �
, ,,
3.B
g f ,
2.B
"
\
I
i
1.8
8
5
18
.5
28
25
38
35
ROI
48
45
58
55
6B
65
?e
75
88
,, ,, 'r-""::OOO�
\
Number
FIG. 5. The effect of increasing the percent threshold on the mean rCMRg•c profile of 18 normal volunteer subjects. The
uppermost and lowermost profiles correspond to 5% and 25% thresholded data, respectively. ROI code: 05, cerebel lum; 10, brainstem; 15, midbrain; 20, basal ganglia; 25, thal amus; 30, hippocampus; 35, lateral temporal cortex; 40, opercular cortex; 45, posterior temporal cortex; 50, medial frontal cortex; 55, lateral frontal cortex; 60, calcarine cortex; 65, cuneus; 70, inferior parietal cortex; 75, paracentral cor tex.
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
Planar Region of Interest FIG. 6. The 25-20% difference annulus, here illustrated sche matically, represents the set of voxels lying outside the 20% threshold but within the 25% threshold.
PERCENT THRESHOLDING
TABLE 6. Prediction of neuropsychological test scores
TABLE 4. Prediction of FlO subject
by subject weights SSFI and SSF2
weights SSFI-SSF4
Data set F05 FI5 F20 F25 F25-F20 RIO R25
A87
SSF!
SSF2
SSF3
SSF4
Data set
99.8 99.8 99.5 99.1 97.0 91. 9 89.4
99.0 99.3 99.2 98.1 93.6 97.5 87.5
97.1 99.0 98. 7 97. 9 90. 6 90. 7 73.0
97. 7 99.0 96.5 90.7 43. 8 94.7 35.8
F05 F lO F15 F20 F25 F25-F20 RIO R25
Values are % variance accounted for. SSFI-SSF4: F lO subject weights (n = 29). Fxx: functionalized image data thresholded at xx%; Rxx: raw image data thresholded at xx%; F25-F20: 25-20% difference an nulus.
jeet weights for the four most significant metabolic co variance patterns (SSF1-SSF4) extracted from the vari ous raw and functionalized data sets were used to predict the subject weights obtained from the functionalized 10% data set (FlO), our "gold standard." A summary of these regression analyses is presented in Table 4. Subject weights derived from the difference annulus and from all of the thresholded data sets except R25 accurately pre dicted the first two subject weights (SSFI and SSF2) from the FlO data set, which were highly correlated with inde pendent measures of disease severity; as anticipated, the FlO metabolic covariance patterns (GISI and GIS2) cor responding to these subject weights were highly corre lated with the metabolic covariance patterns derived from the other thresholded data sets (Table 5). Thus, in our PC4600/FDG data sets, the extraction of covariance pat terns was largely unaffected by either functionalization or thresholding (cf. Strother et aI., 1991). In order to study the effect of thresholding on the strength of specific FDG/PET-neurobehavioral correla tions, subject weights for the two most significant meta bolic covariance patterns derived from combined AIDS normals data sets were used to predict the performance of demented AIDS patients on a battery of neuropsycholog ical tests. The results of a regression analysis using the FlO data set have already been published (Rottenberg et aI., 1987). Predictions obtained from regressing AIDS pa-
TABLE S. Correlation of metabolic covariance patterns derived from the FlO data set with those derived from other thresholded data sets
Data set
GIS1
GIS2
F05 F!5 F20 F25 F25-F20 RIO R25
0.99 0.99 0.99 0.99 0.99 0.99 0.40
0.99 0.99 0.99 0. 98 0.96 0.91 0.43
Values are Pearson's r. GIS!, GIS2: SSM-derived metabolic covariance patterns cor responding to subject weights SSFI and SSF2. Fxx: functionalized image data thresholded at xx%; Rxx: raw image data, thresholded at xx%; F25-F20: 25-20% difference annulus.
VF
TMB
GPD
GPN
72. 3 72.0 71. 1 70. 9 71. 3 72. 3 68.5 68.7
62.0 61.3 59.8 59.8 59. 6 60. 9 55. 3 62.6
76. 4 79. 6 80.0 80.1 80.1 78. 2 83.8 85.5
71.4 74.4 74.8 74.9 74.9 75.3 74.2 80.6
Values are % variance accounted for. VF, verbal fluency; TMB, trailmaking B; GPD and GPN, grooved pegboard dominant and nondominant, respectively. n = 13 patients with AIDS dementia. Fxx: functionalized image data thresholded at xx%; Rxx: raw image data thresholded at xx%; F25-F20: 25-20% difference an nulus.
tients' neuropsychological test scores on their subject weights, derived from variously thresholded raw and functionalized data sets and from the 25-20% difference annulus, are illustrated in Table 6. Our results suggest that subject weights derived from all of the data sets, including R25, are equally good predictors of patient per formance.
DISCUSSION AND CONCLUSION Under some commonly encountered imaging conditions, percent thresholding may increase sen sitivity to regional activation; however, the effect of thresholding is determined, inter alia, by the rela tive magnitude of regional activation and by the specific threshold selected. At the spatial resolution of our PC4600 positron camera (approximately 10 mm FWHM), SSM-derived metabolic covariance patterns are extremely robust and virtually thresh old independent. Moreover, in our FDG data sets, SSM metabolic covariance patterns derived from "peak" regional values and from the 25-20% "dif ference annulus" are highly correlated (r > 0.96, p < 0. 0001). All of these observations bear directly on the im portance of thresholding and, also, on the larger issue of region drawing. How large a region should one draw? In addition, how can one be sure that a given ROI contains one and only one source of co varying metabolic activity? The difference-annulus concept suggests an answer: If the activity of the difference annulus is highly correlated with peak activity, it is unlikely that the ROI contains more than one detectable source of covarying metabolic activity. REFERENCES Anderson NE, Posner JB, Sidtis JJ, Moeller JR, Strother SC, Dhawan V, Rottenberg DA (1988) The metabolic anatomy of
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
D. A. ROTTENBERG ET AL.
A88
paraneoplastic cerebellar degeneration. Ann Neurol23:533540 Evans AC, Diksic M, Yamamoto YL, Kato A, Dagher A, Redies C, Hakim A (1986) Effect of vascular activity in the deter mination of rate constants for the uptake of 18F-labelled 2fluoro-2-deoxy-D-glucose: Error analysis and normal values in older subjects. J Cereb Blood Flow Metab 6:724-738 Eycleshymer AC, Schoemaker DM (1911) A Cross-Sectional Anatomy. New York, Appleton-Century Ginos JZ, French R, Reamer R (1987) The synthesis of high radiochemical purity 2-[18Fl-fluoro-2-deoxy-D-glucose with out the use of preparative HPLC. J Label Comp Radio pharm 24:805-815 Kearfott KJ, Carroll LR (1984) Evaluation of the performance characteristics of the PC4600 positron emission tomography. J Comput Assist Tomogr 8:502-513 Moeller JR, Strother SC, Sidtis n, Rottenberg DA (1987) The scaled subprofile model: A statistical approach to the anal ysis of functional patterns in positron emission tomographic data. J Cereb Blood Flow Metab 7:649--658
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
Phelps ME, Huang SC, Hoffman EJ, Selin C, Sokoloff L, Kuhl DE (1979) Tomographic measurement of local cerebral glu cose metabolic rate in humans with eSF)2-fluoro-2deoxy-D-glucose. Validation of method. Ann Neurol6:3713 88 Rottenberg DA, Moeller JR, Strother SC, Sidtis JJ, Navia BA, Dhawan V, Ginos JZ, Price RW (1987) The metabolic pa thology of the AIDS dementia complex. Ann Neurol22:700706 Strother SC, Allard C, Moeller JR, Sidtis n, Dhawan V, Ginos JZ, Sergi M, Rottenberg DA (1987) Methodological factors affecting patterns of regional cerebral glucose metabolism as determined by 18F-fluorodeoxyglucose/positron emission to mography. J Cereb Blood Flow Metab 7(suppl l):S443 Strother SC, Liow JS. Moeller JR. Sidtis n, Dhawan V, Rotten berg DA (1991) Absolute quantitation in neurological PET: do we need it? J Cereb Blood Flow Metab 11:A3-AI6 Tewson TJ (1983) Synthesis of no-carrier-added fluorine-18 2fluoro-2-deoxyglucose. J Nucl Med 24:718-721
11:A83-A88
©
1991 The International Society of Cerebral Blood Flow and Metabolism
Effects of Percent Thresholding on the Extraction of
eSP]Pluorodeoxyglucose Positron Emission Tomographic Region-of-Interest Data
D. A. Rottenberg, J. R. Moeller, S. C. Strother, V. Dhawan, and M. L. Sergi Department of Neurology, Memorial Sloan-Kettering Cancer Center, New York, New York, U.S.A.
Summary: Although we and others have employed a
determined by a number of factors, including the relative magnitude of regional activation, ROI size, and the spe cific threshold selected. The difference-annulus concept is proposed as a means to study the effects of different region drawing and thresholding strategies, and to deter mine if a given ROI contains one and only one source of covarying metabolic activity. Key Words: Positron emis sion tomography (PET)-Thresholding-Region of in terest (ROI)-Regional cerebral glucose metabolism (rCMRg1J·
thresholding strategy to extract "peak" values from pos itron emission tomographic (PET) regions of interest (ROIs), the effects of peak picking on fitted fluorodeox yglucose rate constants, regional metabolic rate for glu cose (rCMRglc) profiles, patterns of regional metabolic covariation, and PET-neurobehavioral correlations have not been systematically investigated. Our results suggest that under some commonly encountered imaging condi tions percent thresholding may increase sensitivity to re gional activation; however, the effect of thresholding is
We and others have employed a thresholding strategy to extract stable "peak" values from pos itron emission tomographic (PET) regions of inter est (ROls); however, no one, to our knowledge, has systematically investigated the effects of peak picking on fitted fluorodeoxyglucose (FDG) rate constants, regional metabolic rate for glucose (rCMRglc) profiles, patterns of regional metabolic covariation, or PET-neurobehavioral correlations. Strother et al. (1987) previously concluded that thresholding did not affect the distribution of co variance-pattern weights in a small group of normal volunteer subjects.
Address correspondence and reprint requests to Dr. D. A. Rottenberg at PET Imaging Service (lIP), Veterans Administra tion Medical Center, Minneapolis, MN 55417, U. S.A. Abbreviations used: COV, coefficient of variation; FDG, flu orodeoxyglucose; FWHM, full width at half-maximum; GIS, group invariant subprofile; GM, gray matter; PET, positron emission tomography; rCBV, regional cerebral blood volume; ROI, region of interest; SSF, subprofile scaling factor; SSM, scaled subprofile model.
FIG. 1. The effect of variously thresholding raw-count FDG/PET images. Percent thresholds are indicated below and to the left of each image. Note that RDls are drawn with out regard to exact anatomical boundaries and that thresh olded voxels are not always contiguous.
A83
D. A. ROTTENBERG ET AL.
A84
WHAT IS THRESHOLDING?
Thresholding refers to the selection of voxels with numerical values that fall within a predeter mined range, specified by lower and upper bound ary values, e.g., [Vmax
'
(100 - 1)/100]
�
V
�
Vmax
where Vmax refers to the maximum voxel value, v to a thresholded voxel value, and 1 to the upper per cent threshold. Within the context of this discussion, a 5% threshold value for a given ROI signifies that the A
2.88 1.75 ::Jl > U
.... ....
+------t 5
mm
WHY THRESHOLD?
Because of the limited spatial resolution of PET cameras [0.5-1.0 mm full width at half-maximum (FWHM)] relative to the width of the human corti-
FWHM
1.5. +------+ 5
1.50
: , ,
1.25
'
,
B.75
U
U
....
10
mm
Distance
FWHM
1.5e +------t 10
--I II II II I I I I
1.5.
1.25
mm
FWHM
1.25
>
1.••
U
1:5
(l) 0:::
0.25
+------t
.... ....
13.50
8.8e +--"'----'--"---+
Di stcmce
1.75
13.75
(l) >
13.50
2 •••
1.130
-----1
+__"" 8.88 +-_L----'-_�_�_'_____'__________
B
FWHM
::Jl >
1.ee
mm
1.25
---J
1:5
(l) 0:::
upper 5% ofROI voxel values are averaged and that the average value is assigned to the ROI. Figure 1 illustrates the effect of various percent thresholds on raw-count eSF]fluorodeoxyglucose (FDG)/PET images. Note that thresholding permits a relaxation of stringent spatial constraints: ROls can be drawn quickly, without regard to exact anatomical bound aries, and, within a given ROI, thresholded voxels need not be contiguous.
... ....
•• 5.
1:5
(l) 0:::
•.25
•.••
- ---'_---'-L_--'-_-'-_-'-__-=____+ -/""'-_ Distance
FIG.2. The effect of spatial resolution on peak and mean ROI values. Idealized slabs of gray matter in a cortical convolu tion have been imaged with 5 mm (A) or 10 mm (B) [full width at half-maximum (FWHM)] resolution. The underlying corti cal structure (4 mm cortex, 1 mm gap, 4 mm cortex, 5 mm gap, 5 mm cortex with superimposed 2 mm focus of 50% activation) is indicated by dotted lines and the reconstructed activity profile by solid lines.
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
9.50
8.25
e.ee -F------=�--1 Distance FIG. 3. The effect of increasing levels of cortical activati()n c
5 mm (A) and 10 mm (B) FWHM resolution. In each case, th five curves represent (from lowest to highest) 0, 25, 50, 7! and 100% activation. In B, note that the activated right pea lies below the unactivated left peak until the level of actiVe tion reaches 100%.
AS5
PERCENT THRESHOLDING cal ribbon (�3 mm), discrete foci of functional ac tivation are blurred in a more or less Gaussian fash ion. In some situations, thresholding increases sen sitivity to regional activation effects (Figs. 2-4); however, the "blurring" that results from volume averaging with adjacent white matter and/or nonac tivated cortex cannot be removed or minimized by thresholding techniques. The potential benefits to be derived from thresholding depend ultimately on the macroscopic structure of the gray matter (GM) 1.1 +-----+----+---+
A
5
mm
B.9
....
U
mm
spatial resolution
FWHM
loB
....
TABLE 1. Percent thresholding with 5
B.8
(l) >
METHODS AND RESULTS
.... +'
0
B.7
(l) 0::: B.6
B.5 25
5B
75
lBB
Per-cent Ac t i vcd ion B
Effect of thresholding on fitted FDG rate constants
In order to investigate the effect of thresholding on fitted FDG rate constants, 21 serial PET images (10 x 1 min, 5 x 2 min, 3 x 5 min, 3 x 10 min) were acquired with the PC4600 Positron Camera (Kearfott and Carroll, 1984) following the injection of 5-10 mCi of FDG, and the time course of plasma 18F radioactivity was determined by sampling radial arterial blood. Twenty-eight standardized cortical and subcortical ROIs were outlined on recon structed PET brain slices with reference to a neuroana tomical atlas (E ycleshymer and Schoemaker, 1911; Anderson et a!., 1988), and regional GM rate constants (k1-k3) and cerebral blood volume (rCBV) were estimated from the time course of blood and regional brain radio activity (Evans et a!. , 1986). Ten, 15, and 25% thresholds were used to "fix" ROI voxe1s on 45-to-55-min PET
B.8
10
mm
FWHM
B.7 ::JI ....
>
B.7
....
U ....
0
FDG, produced by a modification of Tewson's synthe sis (Tewson, 1983; Ginos et aI. , 1987), was more than 97% radiochemically pure (specific activity of 500 mCilmmol). Twelve demented AIDS patients and 18 normal volunteer subjects were scanned in a standard resting state with eyes patched and with minimal auditory stimulation (Rot tenberg et aI. , 1987; Anderson et aI. , 1988).
B.6
......
(l) 0::: B.5
TABLE 2. Percent thresholding with 10
mm
spatial resolution
B.5 25
5B
75
lB0
Percent activation
Per-cent Ac t 1 va t ion FIG. 4. The effect of percent thresholding at various levels of cortical activation. The left (A) and right (8) panels are based on the data in Tables 1 and 2, respectively. In each case, the dotted line represents the mean regional value (obtained by averaging all reconstructed voxel values within the outlines of the underlying cortical structure), and the solid lines rep resent the variously thresholded ROI values; the ordering of the percent-threshold profiles (i.e., from 5 to 25%) is indi cated in both panels. Note that thresholding increases our ability to detect activation effects at 5 mm resolution but not at 10 mm resolution.
ROI mean 5% THR 10% THR 15% THR 20% THR 25% THR
0%
25%
50%
75%
100%
0.527 0.620 0.611 0.596 0.579 0.541
0.549 0.623 0.612 0.586 0.564 0.557
0.570 0.625 0.602 0.586 0.582 0.577
0.592 0.627 0.612 0.609 0.602 0.596
0.614 0.666 0.643 0.635 0.631 0.622
All values represent relative activity and refer to plots in Fig. 4B. THR: threshold.
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
A86
D. A. ROTTENBERG ET AL. TABLE 3. Effect of percent thresholding on fitted values of krk3' rCBV, and rCMRgic Percent threshold 10% 0.084 0.108 0.082 0.079
k J k2 k3 rCBV
8.58
rCMRglc
±
0.017 0.053 0.031 0.026
(20%) (49%) (38%) (32%)
0.083 0.107 0.083 0.077
±
1.57 (18%)
8.92
± ± ±
25%
15%
±
0.015 0.051 0.028 0.026
(18%) (47%) (34%) (34%)
0.077 0.097 0.078 0.076
±
1.20 (13%)
8.18
± ± ±
±
0.015 0.048 0.030 0.025
±
1.47 (18%)
± ± ±
(19%) (49%) (39%) (33%)
n 18 normal volunteer subjects, age 27 ± 5 years (mean ± SO). Units: k -k3' min-J; rCBV, ml/g; rCMRglc, mg min-J 100 ml-J. J All values are means across subjects ± SO (COV) of the within-subject means across all cortical and subcortical gray-matter regions of interest. rCMRglc Cp/LC . k k3/(k2 + k3)' J =
=
images (frame 21), and these voxels were then "re played" on frames 1-20 in order to define the time course of ROI radioactivity. The results of fitting rate constants and rCBV to thresholded ROI data are illustrated in Table 3. Although there is a trend toward lower coefficients of variation (COVs) with a 15% threshold, these data sug gest that percent thresholding has little effect on the pre cision with which mean GM FDG rate constants can be estimated. Effect of thresholding on mean
rCMRgIc profiles
ROI values and
Effect of thresholding on patterns of regional metabolic covariation
With FDG/PET (and 0-15/PET) images, increasing the percent threshold from 5 to 25% increases the number of nonactivated and/or non-gray-matter voxels being aver aged and decreases the mean ROI value in a systematic fashion. To illustrate this phenomenon, 28 standardized 19.9 9.8 0'
8.8
(S) (S)
7.8
...... C
6.8
E ...... 0" E
ROIs were outlined on reconstructed PET brain slices from the normal volunteer subjects, "functionalized" (i.e., converted from cps/voxel to rCMRglc units) accord ing to Phelps et al. (1979); mean rCMRglc values were then calculated across subjects for 5, 10, 15, 20, and 25% thresholds and displayed in the form of metabolic pro files. As is evident from Fig. 5, changing the percent threshold appears to scale the mean rCMRglc profile with out altering its shape.
Next, we investigated the stability of group invariant subprofiles (GISs), patterns of metabolic covariation ex tracted using the scaled subprofile model (SSM, Moeller et aI., 1987). GISs and their respective subject weights (subprofile scaling factors, SSFs) were derived from FDG/PET data sets containing ROI values for 12 patients with AIDS dementia and 17 normal volunteer subjects (Rottenberg et aI., 1987). Both raw-count (cps/voxel) and functionalized (glucose metabolic rate) data sets were an alyzed. Raw-count ROI data were thresholded at 10 and 25%, and functionalized ROI data were thresholded at 5, 10, 15, 20, and 25%; additionally, a 25-20% "difference annulus" was operationally defined as the set of voxels lying outside the 20% threshold but within the 25% threshold (Fig. 6). Using multiple linear regression, sub-
�.8
/�'�,''.
•. 8 ::J ...... CJ 0:: L U �
, ,,
3.B
g f ,
2.B
"
\
I
i
1.8
8
5
18
.5
28
25
38
35
ROI
48
45
58
55
6B
65
?e
75
88
,, ,, 'r-""::OOO�
\
Number
FIG. 5. The effect of increasing the percent threshold on the mean rCMRg•c profile of 18 normal volunteer subjects. The
uppermost and lowermost profiles correspond to 5% and 25% thresholded data, respectively. ROI code: 05, cerebel lum; 10, brainstem; 15, midbrain; 20, basal ganglia; 25, thal amus; 30, hippocampus; 35, lateral temporal cortex; 40, opercular cortex; 45, posterior temporal cortex; 50, medial frontal cortex; 55, lateral frontal cortex; 60, calcarine cortex; 65, cuneus; 70, inferior parietal cortex; 75, paracentral cor tex.
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
Planar Region of Interest FIG. 6. The 25-20% difference annulus, here illustrated sche matically, represents the set of voxels lying outside the 20% threshold but within the 25% threshold.
PERCENT THRESHOLDING
TABLE 6. Prediction of neuropsychological test scores
TABLE 4. Prediction of FlO subject
by subject weights SSFI and SSF2
weights SSFI-SSF4
Data set F05 FI5 F20 F25 F25-F20 RIO R25
A87
SSF!
SSF2
SSF3
SSF4
Data set
99.8 99.8 99.5 99.1 97.0 91. 9 89.4
99.0 99.3 99.2 98.1 93.6 97.5 87.5
97.1 99.0 98. 7 97. 9 90. 6 90. 7 73.0
97. 7 99.0 96.5 90.7 43. 8 94.7 35.8
F05 F lO F15 F20 F25 F25-F20 RIO R25
Values are % variance accounted for. SSFI-SSF4: F lO subject weights (n = 29). Fxx: functionalized image data thresholded at xx%; Rxx: raw image data thresholded at xx%; F25-F20: 25-20% difference an nulus.
jeet weights for the four most significant metabolic co variance patterns (SSF1-SSF4) extracted from the vari ous raw and functionalized data sets were used to predict the subject weights obtained from the functionalized 10% data set (FlO), our "gold standard." A summary of these regression analyses is presented in Table 4. Subject weights derived from the difference annulus and from all of the thresholded data sets except R25 accurately pre dicted the first two subject weights (SSFI and SSF2) from the FlO data set, which were highly correlated with inde pendent measures of disease severity; as anticipated, the FlO metabolic covariance patterns (GISI and GIS2) cor responding to these subject weights were highly corre lated with the metabolic covariance patterns derived from the other thresholded data sets (Table 5). Thus, in our PC4600/FDG data sets, the extraction of covariance pat terns was largely unaffected by either functionalization or thresholding (cf. Strother et aI., 1991). In order to study the effect of thresholding on the strength of specific FDG/PET-neurobehavioral correla tions, subject weights for the two most significant meta bolic covariance patterns derived from combined AIDS normals data sets were used to predict the performance of demented AIDS patients on a battery of neuropsycholog ical tests. The results of a regression analysis using the FlO data set have already been published (Rottenberg et aI., 1987). Predictions obtained from regressing AIDS pa-
TABLE S. Correlation of metabolic covariance patterns derived from the FlO data set with those derived from other thresholded data sets
Data set
GIS1
GIS2
F05 F!5 F20 F25 F25-F20 RIO R25
0.99 0.99 0.99 0.99 0.99 0.99 0.40
0.99 0.99 0.99 0. 98 0.96 0.91 0.43
Values are Pearson's r. GIS!, GIS2: SSM-derived metabolic covariance patterns cor responding to subject weights SSFI and SSF2. Fxx: functionalized image data thresholded at xx%; Rxx: raw image data, thresholded at xx%; F25-F20: 25-20% difference annulus.
VF
TMB
GPD
GPN
72. 3 72.0 71. 1 70. 9 71. 3 72. 3 68.5 68.7
62.0 61.3 59.8 59.8 59. 6 60. 9 55. 3 62.6
76. 4 79. 6 80.0 80.1 80.1 78. 2 83.8 85.5
71.4 74.4 74.8 74.9 74.9 75.3 74.2 80.6
Values are % variance accounted for. VF, verbal fluency; TMB, trailmaking B; GPD and GPN, grooved pegboard dominant and nondominant, respectively. n = 13 patients with AIDS dementia. Fxx: functionalized image data thresholded at xx%; Rxx: raw image data thresholded at xx%; F25-F20: 25-20% difference an nulus.
tients' neuropsychological test scores on their subject weights, derived from variously thresholded raw and functionalized data sets and from the 25-20% difference annulus, are illustrated in Table 6. Our results suggest that subject weights derived from all of the data sets, including R25, are equally good predictors of patient per formance.
DISCUSSION AND CONCLUSION Under some commonly encountered imaging conditions, percent thresholding may increase sen sitivity to regional activation; however, the effect of thresholding is determined, inter alia, by the rela tive magnitude of regional activation and by the specific threshold selected. At the spatial resolution of our PC4600 positron camera (approximately 10 mm FWHM), SSM-derived metabolic covariance patterns are extremely robust and virtually thresh old independent. Moreover, in our FDG data sets, SSM metabolic covariance patterns derived from "peak" regional values and from the 25-20% "dif ference annulus" are highly correlated (r > 0.96, p < 0. 0001). All of these observations bear directly on the im portance of thresholding and, also, on the larger issue of region drawing. How large a region should one draw? In addition, how can one be sure that a given ROI contains one and only one source of co varying metabolic activity? The difference-annulus concept suggests an answer: If the activity of the difference annulus is highly correlated with peak activity, it is unlikely that the ROI contains more than one detectable source of covarying metabolic activity. REFERENCES Anderson NE, Posner JB, Sidtis JJ, Moeller JR, Strother SC, Dhawan V, Rottenberg DA (1988) The metabolic anatomy of
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
D. A. ROTTENBERG ET AL.
A88
paraneoplastic cerebellar degeneration. Ann Neurol23:533540 Evans AC, Diksic M, Yamamoto YL, Kato A, Dagher A, Redies C, Hakim A (1986) Effect of vascular activity in the deter mination of rate constants for the uptake of 18F-labelled 2fluoro-2-deoxy-D-glucose: Error analysis and normal values in older subjects. J Cereb Blood Flow Metab 6:724-738 Eycleshymer AC, Schoemaker DM (1911) A Cross-Sectional Anatomy. New York, Appleton-Century Ginos JZ, French R, Reamer R (1987) The synthesis of high radiochemical purity 2-[18Fl-fluoro-2-deoxy-D-glucose with out the use of preparative HPLC. J Label Comp Radio pharm 24:805-815 Kearfott KJ, Carroll LR (1984) Evaluation of the performance characteristics of the PC4600 positron emission tomography. J Comput Assist Tomogr 8:502-513 Moeller JR, Strother SC, Sidtis n, Rottenberg DA (1987) The scaled subprofile model: A statistical approach to the anal ysis of functional patterns in positron emission tomographic data. J Cereb Blood Flow Metab 7:649--658
J Cereb Blood Flow Metab, Vol.
11,
Suppl.
1, 1991
Phelps ME, Huang SC, Hoffman EJ, Selin C, Sokoloff L, Kuhl DE (1979) Tomographic measurement of local cerebral glu cose metabolic rate in humans with eSF)2-fluoro-2deoxy-D-glucose. Validation of method. Ann Neurol6:3713 88 Rottenberg DA, Moeller JR, Strother SC, Sidtis JJ, Navia BA, Dhawan V, Ginos JZ, Price RW (1987) The metabolic pa thology of the AIDS dementia complex. Ann Neurol22:700706 Strother SC, Allard C, Moeller JR, Sidtis n, Dhawan V, Ginos JZ, Sergi M, Rottenberg DA (1987) Methodological factors affecting patterns of regional cerebral glucose metabolism as determined by 18F-fluorodeoxyglucose/positron emission to mography. J Cereb Blood Flow Metab 7(suppl l):S443 Strother SC, Liow JS. Moeller JR. Sidtis n, Dhawan V, Rotten berg DA (1991) Absolute quantitation in neurological PET: do we need it? J Cereb Blood Flow Metab 11:A3-AI6 Tewson TJ (1983) Synthesis of no-carrier-added fluorine-18 2fluoro-2-deoxyglucose. J Nucl Med 24:718-721
