[Skip to Content]
Access to paid content on this site is currently suspended due to excessive activity being detected from your IP address 184.73.122.162. Please contact the publisher to request reinstatement.
Sign In
Individual Sign In
Create an Account
Institutional Sign In
OpenAthens Shibboleth
[Skip to Content Landing]
Download PDF
Figure 1.
The locations of the 9 volumes of interest used for the magnetic resonance imaging and magnetic resonance imaging–guided single-photon emission computed tomography data: rostral anterior cingulate (blue), caudal anterior cingulate (green), and posterior cingulate (red) (A); temporal horn (purple), hippocampus (blue), and entorhinal cortex (orange) (B); basal forebrain (blue) and amygdala (yellow) (C); and the banks of the superior temporal sulcus (green) (D).

The locations of the 9 volumes of interest used for the magnetic resonance imaging and magnetic resonance imaging–guided single-photon emission computed tomography data: rostral anterior cingulate (blue), caudal anterior cingulate (green), and posterior cingulate (red) (A); temporal horn (purple), hippocampus (blue), and entorhinal cortex (orange) (B); basal forebrain (blue) and amygdala (yellow) (C); and the banks of the superior temporal sulcus (green) (D).

Figure 2.
An outline of the procedure for estimating the magnetic resonance imaging (MRI)–guided single-photon emission computed tomography (SPECT) measures, including corrections for scatter, nonuniform attenuation, and variable collimator response. OSEM indicates ordered subset expectation maximization.

An outline of the procedure for estimating the magnetic resonance imaging (MRI)–guided single-photon emission computed tomography (SPECT) measures, including corrections for scatter, nonuniform attenuation, and variable collimator response. OSEM indicates ordered subset expectation maximization.

Figure 3.
Correlated receiver operating characteristic curves for discrimination between controls and converters using single-photon emission computed tomography (SPECT) data and magnetic resonance imaging (MRI) data or MRI data and combined SPECT and MRI data. Az indicates area under the receiver operating characteristic curve (given as mean ± SD).

Correlated receiver operating characteristic curves for discrimination between controls and converters using single-photon emission computed tomography (SPECT) data and magnetic resonance imaging (MRI) data or MRI data and combined SPECT and MRI data. Az indicates area under the receiver operating characteristic curve (given as mean ± SD).

Figure 4.
Correlated receiver operating characteristic curves for discrimination between controls and questionables using single-photon emission computed tomography (SPECT) data and magnetic resonance imaging (MRI) data or MRI data and combined SPECT and MRI data. Az indicates area under the receiver operating characteristic curve (given as mean ± SD).

Correlated receiver operating characteristic curves for discrimination between controls and questionables using single-photon emission computed tomography (SPECT) data and magnetic resonance imaging (MRI) data or MRI data and combined SPECT and MRI data. Az indicates area under the receiver operating characteristic curve (given as mean ± SD).

Figure 5.
Correlated receiver operating characteristic curves for discrimination between questionables and converters using single-photon emission computed tomography (SPECT) data and magnetic resonance imaging (MRI) data or MRI data and combined SPECT and MRI data. Az indicates area under the receiver operating characteristic curve (given as mean ± SD).

Correlated receiver operating characteristic curves for discrimination between questionables and converters using single-photon emission computed tomography (SPECT) data and magnetic resonance imaging (MRI) data or MRI data and combined SPECT and MRI data. Az indicates area under the receiver operating characteristic curve (given as mean ± SD).

1.
Jagust  WJBudinger  TFReed  BR The diagnosis of dementia with single photon emission computed tomography. Arch Neurol.1987;44:258-262.
PubMed
2.
Johnson  KADavis  KRBuonanno  FSBrady  TJGrowdon  JH Comparison of magnetic resonance and x-ray computed tomography in dementia. Arch Neurol.1987;44:1075-1080.
PubMed
3.
DeKosky  SShih  WJScmitt  FCoupal  JKirkpatrick  C Assessing utility of single photon emission computed tomography (SPECT) scan in Alzheimer's disease: correlation with cognitive severity. Alzheimer Dis Assoc Disord.1990;4:14-23.
PubMed
4.
Holman  BLJohnson  KAGerada  BCarvalho  PASatlin  A The scintigraphic appearance of Alzheimer's disease: a prospective study using technetium-99m-HMPAO SPECT. J Nucl Med.1992;33:181-185.
PubMed
5.
Johnson  KJones  KHolman  BL  et al Preclinical prediction of Alzheimer's disease using SPECT. Neurology.1998;50:1563-1571.
PubMed
6.
Minoshima  SGiordani  BBerent  SFrey  KAFoster  NLKuhl  DE Metabolic reduction in the posterior cingulate cortex in very early Alzheimer's disease. Ann Neurol.1997;42:85-94.
PubMed
7.
Okamura  NShinkawa  MArai  H  et al Prediction of progression in patients with mild cognitive impairment using IMP-SPECT. Nippon Ronen Igakkai Zasshi.2000;37:974-978.
PubMed
8.
Du  ASchuff  NAmend  D  et al MRI of entorhinal cortex and hippocampus in mild cognitive impairment and Alzheimer's disease. J Neurol Neurosurg Psychiatry.2001;71:441-447.
PubMed
9.
Convit  Ade Leon  MJTarshish  C  et al Specific hippocampal volume reductions in individuals at risk for Alzheimer's disease. Neurobiol Aging.1997;18:131-138.
PubMed
10.
De Toledo-Morrell  LGoncharova  IDickerson  BWilson  RSBennett  DA From healthy aging to early Alzheimer's disease: in vivo detection of entorhinal cortex atrophy. Ann N Y Acad Sci.2000;911:240-253.
PubMed
11.
Killiany  RJGomez-Isla  TMoss  MB  et al Use of structural magnetic resonance imaging to predict who will get Alzheimer's disease. Ann Neurol.2000;47:430-439.
PubMed
12.
Kaye  JASwihart  THowieson  D  et al Volume loss of the hippocampus and temporal lobe in healthy elderly persons destined to develop dementia. Neurology.1997;48:1297-1304.
PubMed
13.
Xu  YJack Jr  CRO'Brien  PC  et al Usefulness of MRI measures of entorhinal cortex versus hippocampus in AD. Neurology.2000;54:1760-1767.
PubMed
14.
Albert  MSMoss  MBTanzi  RJones  K Preclinical prediction of AD using neuropsychological tests. J Int Neuropsychol Soc.2001;7:631-639.
PubMed
15.
Daly  EZaitchik  DCopeland  MSchmahmann  JGunther  JAlbert  MS Predicting conversion to Alzheimer disease using standardized clinical information. Arch Neurol.2000;57:675-680.
PubMed
16.
Hughes  WJBerg  LDanziger  WLCoben  LAMartin  RL A new clinical scale for the staging of dementia. Br J Psychiatry.1982;140:566-572.
PubMed
17.
Folstein  MFolstein  SMcHugh  P "Mini-Mental State": a practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res.1975;12:189-198.
PubMed
18.
McKhann  GDrachman  DFolstein  MKatzman  RPrice  DStadlan  EM Clinical diagnosis of Alzheimer's disease: report of the NINCDS-ADRDA Work Group under the auspices of Department of Health and Human Services Task Force on Alzheimer's disease. Neurology.1984;34:939-944.
PubMed
19.
National Institute on Aging, Reagan Institute Working Group on Diagnostic Criteria for the Neuropathological Assessment of Alzheimer's Disease Consensus recommendations for the postmortem diagnosis of Alzheimer's disease. Neurobiol Aging.1997;18(suppl):S1-S2.
PubMed
20.
Killiany  RJMoss  MBAlbert  MSSandor  TJolesz  F Temporal lobe regions on magnetic resonance imaging identify patients with early Alzheimer's disease. Arch Neurol.1993;50:949-954.
PubMed
21.
Genna  SSmith  AP The development of ASPECT, an annular single crystal brain camera for high efficiency SPECT. IEEE Trans Nucl Sci.1988;35:654-658.
22.
El Fakhri  GMoore  SCMaksud  PAurengo  AKijewski  MF Absolute activity quantitation in simultaneous I-123/Tc-99m brain SPECT. J Nucl Med.2001;42:300-308.
PubMed
23.
El Fakhri  GKijewski  MFMoore  SC Absolute activity quantitation from projections using an analytical approach: comparison with iterative methods in Tc-99m and I-123 brain SPECT. IEEE Trans Nucl Sci.2001;48:768-773.
24.
Moore  SCKijewski  MFMuller  SPRybicki  FZimmerman  RE Evaluation of scatter compensation methods by their effects on parameter estimation from SPECT projections. Med Phys.2001;28:278-287.
PubMed
25.
Shepp  LAVardi  Y Maximum likelihood reconstruction for emission tomography. IEEE Trans Med Imaging.1982;1:113-122.
26.
Hudson  HMLarkin  RS Accelerated image reconstruction using ordered subsets of projection data. IEEE Trans Med Imaging.1994;13:601-609.
27.
Pelizzari  CAChen  GTYSpelbring  DRWeichselbaum  RRChen  CT Accurate three-dimensional registration of CT, PET, and/or MR images of the brain. J Comput Assist Tomogr.1989;13:20-26.
PubMed
28.
Holman  BLZimmerman  REJohnson  KA  et al Computer-assisted superimposition of magnetic resonance and high-resolution technetium-99m-HMPAO and thallium-201 SPECT images of the brain. J Nucl Med.1991;32:1478-1484.
PubMed
29.
Rencher  AC Multivariate Statistical Inference and Applications.  New York, NY: John Wiley & Sons Inc; 1998.
30.
Metz  CE ROC methodology in radiologic imaging. Invest Radiol.1986;21:720-733.
PubMed
31.
El Fakhri  GBuvat  IBenali  HTodd-Pokropek  ADi Paola  R Relative impact of scatter, collimator response, attenuation, and finite spatial resolution corrections in cardiac SPECT. J Nucl Med.2000;41:1400-1408.
PubMed
32.
Albert  MSMoss  MB Neuropsychological approach to preclinical identification of Alzheimer's disease.  In: Squire  L, Schacter  D, eds. Neuropsychology of Memory. New York, NY: Guilford Publications; 2002:248-262.
Original Contribution
August 2003

MRI-Guided SPECT Perfusion Measures and Volumetric MRI in Prodromal Alzheimer Disease

Author Affiliations

From the Department of Radiology, Brigham and Women's Hospital and Harvard Medical School, Boston, Mass (Drs Fakhri and Kijewski and Messrs Syrkin, Becker, and Zimmerman); the Departments of Radiology (Dr Johnson), Neurology (Drs Johnson and Albert), and Psychiatry (Dr Albert), Massachusetts General Hospital and Harvard Medical School, Boston; and the Departments of Anatomy and Neurobiology, Boston University Medical Center (Dr Killiany).

Arch Neurol. 2003;60(8):1066-1072. doi:10.1001/archneur.60.8.1066
Abstract

Objective  To identify group differences in the prodromal phase of Alzheimer disease (AD) using quantitative single-photon emission computed tomography (SPECT) perfusion and magnetic resonance imaging (MRI) volume measures within specific volumes of interest.

Setting  Gerontology research unit.

Participants  There were 17 healthy controls, 56 nondemented patients with memory problems who did not develop AD during 3 to 5 years of follow-up (questionables), and 27 nondemented patients with memory problems who developed AD during follow-up (converters).

Methods  A Tc 99m hexamethylpropyleneamine oxime SPECT study and an MRI were performed in each participant at baseline. Mean SPECT activity concentration and MRI volume were estimated within 9 structures: rostral anterior cingulate, caudal anterior cingulate, posterior cingulate, hippocampus, entorhinal cortex, basal forebrain, temporal horn, amygdala, and the banks of the superior temporal sulcus. Data were analyzed using overall and pairwise discriminant analysis, and performance in pairwise group discrimination was measured using correlated receiver operating characteristic curve analysis.

Results  The overall (3-group) discriminant function was significant for SPECT (F test, P<.001) and MRI (F test, P<.0001). For the SPECT analysis, the ranking of structures for discriminating among the 3 groups was, in order of decreasing discriminating power, caudal anterior cingulate, temporal horn, superior temporal sulcus, entorhinal cortex, hippocampus, rostral anterior cingulate, amygdala, basal forebrain, and posterior cingulate. For the MRI analysis, this ranking was entorhinal cortex, superior temporal sulcus, temporal horn, hippocampus, amygdala, caudal anterior cingulate, rostral anterior cingulate, basal forebrain, and posterior cingulate. Combining the 2 modalities yielded significantly better discrimination performance than did either alone. Furthermore, the correlation between SPECT and MRI measures was low.

Conclusion  Measures of structure activity concentration and volume carry independent information; both reveal group differences in prodromal AD.

SINGLE-PHOTON EMISSION computed tomography (SPECT) reveals perfusion abnormalities in patients with established Alzheimer disease (AD).14 The most consistent finding reported in these studies is decreased perfusion in the temporoparietal association neocortex in mildly and moderately impaired patients with probable AD compared with healthy controls. More recently, several research groups have attempted to identify brain perfusion patterns that predict subsequent development of AD. These efforts have practical, as well as theoretical, significance, because early prediction of AD would make it possible to implement strategies to prevent or delay dementia. Johnson et al5 used principal component analysis to identify decreased perfusion in the hippocampal-amygdaloid complex and in the anterior and posterior cingulate in prodromal AD. This approach does not require a priori assumptions about the locations of discriminating regions, but it may not yield insight into the role of particular brain structures in the development of AD. Other SPECT studies6,7 targeted specific volumes of interest (VOI) and reported that perfusion in the posterior cingulate declines in prodromal AD. None of these studies targeted all of the small brain regions believed to be involved in prodromal AD. Several magnetic resonance imaging (MRI) studies813 have, similarly, sought to determine whether decreased volume in certain brain structures characterizes prodromal AD. Significant volume changes in the entorhinal cortex and the hippocampus are the most commonly reported finding.813

In the present study, we compared MRI volume estimates with SPECT estimates of activity concentration in MRI-guided VOI. The goals were to determine the accuracy of each modality alone in identifying group differences in the prodromal phase of AD and to determine whether the combination of these modalities is significantly better than either one taken separately.

METHODS
PARTICIPANTS

One hundred individuals participated in the study after providing informed consent according to institutional guidelines. These individuals were participants in a large longitudinal study of prodromal AD.14,15 To be included in the study, participants had to be 65 years of age or older and free of significant underlying medical, neurologic, or psychiatric illness based on standard laboratory test results and clinical evaluation findings. Furthermore, they had to be normal or questionable by the Clinical Dementia Rating (CDR) criterion,16 which stages individuals according to their functional ability, with 0 representing normal function and 5 representing the terminal phase of dementia. No one was excluded on the basis of sex or racial or ethnic identity.

At baseline, 17 participants had normal cognition (CDR, 0.0) and 83 met the criteria for "questionable AD" (CDR, 0.5). The mean ± SD ages of the 2 groups were nearly equal (73.8 ± 4.8 and 72.9 ± 6.1 years), as were the mean ± SD Mini-Mental State Examination scores17 (29.4 ± 0.9 and 29.1 ± 1.3). After enrollment in the study, participants were evaluated annually. After 3 to 5 years of follow-up, participants were divided into 3 groups—controls, questionables, and converters—based on their functional status at baseline and at follow-up. (Additional details are available in the study by Albert et al.14)

Controls

This group consisted of 17 individuals (mean ± SD age, 73.8 ± 4.8 years) who entered the study with normal cognition (CDR, 0.0; mean ± SD Mini-Mental State Examination score, 29.4 ± 0.9) and who remained cognitively intact after 3 to 5 years of annual follow-up evaluations (CDR, 0.0; mean Mini-Mental State Examination score, 29.3).

Questionables

This group consisted of 56 individuals (mean ± SD age, 72.5 ± 6.9 years). At baseline, these individuals were not demented but had evidence of memory impairment in daily life (CDR, 0.5; mean sum of boxes, 1.01). After 3 to 5 years of follow-up, these individuals remained nondemented but continued to have cognitive impairments (CDR, 0.5; mean sum of boxes, 1.51).

Converters

This group consisted of 27 individuals (mean ± SD age, 73.7 ± 4.1 years). At baseline, these individuals were nondemented but had evidence of mild memory impairment (CDR, 0.5; mean sum of boxes, 1.36). After 3 to 5 years of follow-up, their cognitive difficulties had progressed to the point where they met National Institute of Neurological and Communicative Disorders and Stroke/Alzheimer's Disease and Related Disorders Association criteria for probable AD (CDR, 1.1; mean sum of boxes, 5.25).18 The annual medical, neurologic, psychiatric, and laboratory evaluations were augmented, as needed, to exclude other potential causes of cognitive decline to establish the diagnosis of probable AD. The SPECT data were not provided to the clinicians making the diagnosis. Four participants died, and a diagnosis of definite AD was confirmed by autopsy examination.19

IMAGING PROCEDURES

At baseline, all study participants underwent SPECT and MRI. Although the imaging data were obtained at baseline, the data were analyzed using the participant's status on follow-up.

MRI Acquisitions and Procedures

Each participant underwent T1-weighted gradient echo MRI (1.5T Signa; General Electric Medical Systems, Milwaukee, Wis) of the brain (repetition time, 35 milliseconds; echo time, 5 milliseconds; field of view, 220; flip angle, 45°; slice thickness, 1.5 mm; and matrix size, 256 × 256). Nine VOI were outlined manually by skilled operators using methods that have been shown to have high reliability.11,20 Most VOI were selected on the basis of neuropathologic or functional neuroimaging data indicating that they are altered early in the course of AD. However, some of the VOI were selected on the basis of probable involvement in the later stages of disease. Each VOI consisted of a left and right pair of structures.

The brain structures considered in these analyses included the entorhinal cortex (ento), the basal forebrain (basfb), the hippocampus (hipp), the amygdala (amy), the temporal horn (thorn), the banks of the superior temporal sulcus (sts), and 3 sections of the cingulate gyrus: the rostral portion of the anterior cingulate (acing), the caudal portion of the anterior cingulate (mcing), and the posterior cingulate (pcing). Each VOI volume estimate was adjusted for total brain volume, as described in the "Statistical Analysis" subsection. The locations of the MRI VOI are shown in Figure 1.

SPECT Acquisitions and Procedures

Brain SPECT was performed using a dedicated brain gamma camera (CeraSPECT; Digital Scintigraphics Inc, Waltham, Mass) with stationary annular thallium-activated sodium iodide crystal, within which rotates a collimator consisting of 3 parallel-hole segments. Intrinsic spatial resolution is 3.6-mm full width at half maximum, and the system spatial resolution is 8.2-mm full width at half maximum at the center of the field of view21 for technetium (99mTc) (140 keV). Planar views (projections) were acquired 20 minutes after injection of a mean ± SD of 740.0 ± 37.0 MBq (20.0 ± 1.0 mCi) of 99mTc hexamethylpropyleneamine oxime (Ceretec; Amersham, Buckinghamshire, England) with the participants supine, at rest, with eyes open in a darkened room with ambient noise. One hundred twenty projections (128 × 64) were acquired in 30 minutes (isotropic pixel dimension, 1.67 mm) in 13 energy windows encompassing the 80- to 154-keV energy range.

The same 9 VOI were evaluated for SPECT as for MRI. Activity estimation in the SPECT VOI was performed after correcting for scatter, nonuniform attenuation, and variable collimator response using a quantitation strategy based on recent work that has been validated by Monte Carlo simulations22 and by phantom studies.23 First, acquired projections were corrected for scatter using a general spectral method24 by which scatter-corrected projections are formed by linear combination of the pixel counts in each energy window. In a previous study,24 the weighting scheme for data within seventeen 4-keV energy windows, ranging from 92 to 160 keV, was optimized on the basis of the accuracy and precision with which lesion and background activity could be simultaneously estimated. Because the data in the present study were collected in different energy windows, weights for these windows were determined from the original weights by using spline interpolation.23

After correction for scatter, the projections were reconstructed by the maximum likelihood expectation maximization approach25 using an accelerated algorithm based on use of ordered subsets (OSEM)26 to yield the perfusion image.

We compensated for the variable collimator response by modeling the distance-dependent component of spatial resolution in the iterative OSEM algorithm.26 Attenuation correction was also performed directly in the reconstruction algorithm by modeling the nonuniform attenuation distribution in the OSEM algorithm.23 The attenuation map was estimated individually for each participant by using the following procedure. First, projections were reconstructed using OSEM (6 subsets and 8 iterations) without corrections for attenuation and collimator response. This preliminary SPECT image was registered to the MRI using a brain surface–based rigid body transformation.27,28 The MRI volume was segmented into bone and soft tissue compartments to yield an attenuation map that was used to correct for nonuniform attenuation as previously described.23 Finally, a second OSEM reconstruction (6 subsets) was performed, with corrections for attenuation and variable collimator response incorporated into the iterative algorithm; reconstructed volumes corresponding to 9, 11, 13, 15, 17, and 19 iterations were saved. Mean regional perfusion was calculated from the 3-dimensional volume within the structure boundaries defined on the registered MRI. Figure 2 illustrates the steps involved in the correction and quantitation of the MRI-guided SPECT measures.

STATISTICAL ANALYSIS
Overall Linear Discriminant Analysis

An overall discriminant analysis29 was performed to determine whether the 9 VOI values together significantly differentiated the 3 groups (controls, questionables, and converters). For the SPECT analysis, 10 variables were used: the 9 structure activity concentration estimates and the total brain SPECT activity, included to adjust for any possible differences between the groups based on this variable. The covariate method of adjustment for total brain SPECT activity was used in these analyses because it allows a correction to occur only when there is significant correlation between total perfusion and the activity concentration in a particular brain structure. An analogous overall discriminant analysis was performed for the MRI data; the 10 variables included the 9 volume estimates and a measure of intracranial volume. The significance of each discriminant function was tested using an F approximation to Wilks Λ.29 The statistical power of the overall analysis is based on the overall sample size, that is, 100 patients, rather than on the sizes of the individual groups.

Pairwise Discriminant Analysis

Stepwise discriminant analyses were performed to identify the structures that were most effective in discriminating between the pairs of groups: controls vs converters, controls vs questionables, and questionables vs converters. For each binary analysis, the most discriminating variable was selected by comparing values of Wilks Λ for the 9 structures of interest (excluding total activity or volume). The next variable was selected by maximizing the partial F statistic for adding a second variable to the first. This procedure was repeated, maximizing the partial F statistic for adding another variable to the set already chosen, until all 9 structures had been ranked. This ranking procedure was performed for 3 sets of variables: (1) the 9 SPECT structure activity estimates, (2) the 9 MRI structure volume estimates, and (3) the combination of SPECT and MRI measures (the ranking procedure was terminated after 9 of the 18 variables had been chosen). The stepwise discriminant analyses were repeated on 2 subgroups of patients obtained by dividing each patient group into 2 to assess the robustness of the rankings (n = 49 and 51).

For each pairwise analysis, discriminant scores were calculated for each participant based on the discriminant function for all 9 variables. Next, a correlated receiver operating characteristic (ROC) curve was fitted to the discriminant score data for each analysis.30 The ROC curve is a plot of true-positive rate (sensitivity) vs false-positive rate (1 − specificity) as the decision threshold is varied. Performance in the binary discrimination tasks is specified by the area under the ROC curve (Az); an Az of 1.0 corresponds to perfect performance, and an Az of 0.5 implies chance performance. Correlated ROC curves were computed for the SPECT and the MRI measures and for the MRI and the combined SPECT and MRI performance. The ROC curve-fitting procedure provided estimates of correlation between the MRI and the SPECT data and between the MRI and the combined SPECT and MRI data.

RESULTS
OVERALL LINEAR DISCRIMINANT ANALYSIS

The overall discriminant function was significant for SPECT (F test, P<.001) and MRI (F test, P<.0001). The most significant overall discriminant function among the 3 groups of patients was obtained for 13 iterations; therefore, the SPECT activity estimates from images reconstructed by 13 iterations of OSEM were used for the remaining analyses. For the SPECT data, the ranking of structures for discriminating among the 3 groups was, in order of decreasing discriminating power, mcing, thorn, sts, ento, hipp, and acing. For the MRI analysis, this ranking was ento, sts, thorn, hipp, amy, and mcing.

PAIRWISE DISCRIMINANT ANALYSES
Controls vs Converters

Analysis of the SPECT data revealed that the structure with the most power to discriminate controls and converters was mcing. The structure that, combined with mcing, had the next most discriminating power was pcing (it is possible that 2 other structures together have more discriminating power than do mcing and pcing combined; this is also true of later-ranking results). The 6 structures with the most discriminating power were mcing, pcing, sts, basfb, acing, and amy. For the MRI analysis, this ranking was ento, thorn, amy, basfb, sts, and acing. When the SPECT and MRI measures were combined, the best structures discriminating between controls and converters (of 18 structures) included MRI and SPECT variables: ento_MRI, mcing_SPECT, mcing_MRI, thorn_MRI, acing_MRI, and sts_SPECT. Similar ranking results were obtained using the subgroups of patients.

Controls vs Questionables

The ranking of the best SPECT variables was, in order of decreasing discriminating power, thorn, mcing, basfb, hipp, pcing, and sts. For the MRI analysis, this ranking was ento, amy, sts, basfb, mcing, and pcing. When the SPECT and MRI measures were combined, the best structures discriminating between controls and questionables were ento_MRI, amy_MRI, sts_MRI, basfb_MRI, thorn_SPECT, and sts_SPECT. The best 6 structures resulting from the pairwise discriminant analyses of the 2 subgroups of patients yielded similar results for the 4 best structures.

Questionables vs Converters

The ranking of the 6 best SPECT variables was, in order of decreasing discriminating power, amy, sts, basfb, mcing, hipp, and acing. For the MRI analysis, this ranking was acing, sts, basfb, thorn, amy, and hipp. When the SPECT and MRI variables were combined, the best structures discriminating between questionables and converters were amy_SPECT, acing_MRI, basfb_MRI, basfb_SPECT, pcing_SPECT, and pcing_MRI. These results were consistent with those obtained for each subgroup of patients.

Figure 34, and 5 show the ROC curves measuring the performance in discriminating between controls and converters, controls and questionables, and questionables and converters, respectively, when using SPECT, MRI, or combined SPECT and MRI data. For SPECT, the mean ± SD Az was 0.962 ± 0.028 for controls vs converters, 0.969 ± 0.019 for controls vs questionables, and 0.879 ± 0.037 for questionables vs converters. For MRI, the comparable mean ± SD Az values were 0.990 ± 0.010, 0.986 ± 0.010, and 0.843 ± 0.049, respectively.

The mean ± SD Az values were significantly higher for combined SPECT and MRI data than for each modality alone: 0.995 ± 0.010 for controls vs converters, 0.987 ± 0.013 for controls vs questionables, and 0.935 ± 0.026 for questionables vs converters (P<.05). The correlation coefficients between the SPECT and the MRI data for the binary discrimination tasks were low (0.0104 for controls vs converters, 0.0083 for controls vs questionables, and 0.0229 for converters vs questionables). The correlation coefficients between the SPECT and the combined MRI and SPECT data were 0.4585, 0.2556, and 0.3031, respectively; these values are higher because the SPECT data contributed to both ROC curves.

COMMENT

The SPECT and MRI variables obtained at baseline are highly significant predictors of the future development of AD. Although MRI and SPECT measures yielded similar discrimination accuracy (Az = 0.843 and 0.879, respectively) in the task of greatest clinical interest, that is, discrimination between questionables and converters, closer examination of these curves shows that MRI yielded better performance at the clinically interesting low false-positive rate of less than 10% (high specificity regime), whereas SPECT outperformed MRI at false-positive rates greater than 10%. The MRI measures yielded greater discrimination for controls vs converters and controls vs questionables than did the SPECT measures. The higher performance obtained in this work compared with previous studies11 resulted from the larger number of brain structures considered (9 instead of 3) and the inclusion of structures that contribute significantly to the discrimination, for example, the basal forebrain, the hippocampus, the temporal horn, the amygdala, and the banks of the superior temporal sulcus.

Most important, combining the best MRI and SPECT measures yielded systematically better results than using either MRI or SPECT data alone, implying that independent information is present in the SPECT and MRI data. This is also seen in the ranking of the structures, which differed for the 2 modalities, and in the ranking obtained when combining the best SPECT and MRI measures as perfusion and volume of certain structures, such as mcing, were both selected. Finally, the low correlations (0.0083-0.0229) between data from the 2 modalities further confirm the independent character of the information present in SPECT and MRI. This suggests that the 2 modalities, used in a complementary manner, will yield the best information on prodromal AD.

Even when combining SPECT and MRI information, the performance in discriminating between questionables and converters, the most clinically interesting task, was significantly less (Az = 0.93) than that for discrimination between controls and converters (Az = 0.99) or controls and questionables (Az = 0.98) (P<.01). This is most likely because some of the questionable participants are destined to develop AD at a later time and, therefore, have SPECT perfusion and MRI volume abnormalities that are similar to those of converters. The decline in sum of boxes among the participants who remained questionable on follow-up (from 1.01 to 1.51) supports this hypothesis.

It is unlikely that the discriminating power of the SPECT variables can be attributed to disease-related atrophy combined with limited spatial resolution. Our reconstruction methods led to improved spatial resolution and, consequently, minimal confounding between SPECT activity concentration estimates and structure volume.23 To support this assertion, we estimated the effects of limited spatial resolution on estimates of activity concentration within the entorhinal cortex, in which, because of its small size, atrophy effects would be greatest. To estimate the magnitude of the underestimation of entorhinal activity due to this phenomenon, we modeled the entorhinal cortex as a cylinder, with an axial length twice its diameter. The mean diameter of the entorhinal cortex in the present study was 8 mm. The spatial resolution in our images was 6-mm full width at half maximum. This implies an underestimation of entorhinal activity of 12%, taking into account surrounding brain structures. This degree of bias is comparable to the best accuracy (10% error) that can be achieved when compensating for all physical factors affecting quantitative SPECT.31

The specific ranking of the structures is also of interest. For the MRI variables, the most discriminating measures pertain to the medial temporal lobe (eg, the entorhinal cortex). This is consistent with several previous studies8,10 and with the fact that a memory deficit is the initial symptom seen in individuals with prodromal AD. (Additional details are available in the review by Albert and Moss.32) For the SPECT variables, the most discriminating measures pertain to the cingulate, particularly the caudal portion of the anterior cingulate. The middle portion of the cingulate has only recently been identified as demonstrating altered perfusion in prodromal AD and has been hypothesized to be related to the deficits in executive function demonstrated in this early phase of disease.5 Differences between the present findings and those of previous studies are probably related to the large quantity of VOI examined in the present study and to the ranking strategy.

Moreover, the estimates of the Azs are affected by resubstitution bias, since the binary classification tasks were accomplished using rules established with the same participants being classified. Therefore, the ROC curve performance measures obtained herein may be better than those achieved in a new group of individuals. They can, however, be viewed as upper bounds on the ability of the image-derived measures to discriminate the groups.

In conclusion, quantitative brain SPECT and MRI identify group differences in the prodromal phase of AD. The most discriminating brain regions identified by SPECT are task dependent and differ from those identified by MRI. Combining information from SPECT and MRI yields the best discrimination performance.

Back to top
Article Information

Corresponding author and reprints: Georges El Fakhri, PhD, Department of Radiology, Brigham and Women's Hospital, Boston, MA 02115 (e-mail: elfakhri@bwh .harvard.edu).

Accepted for publication March 11, 2003.

Author contributions: Study concept and design (Drs El Fakhri, Kijewski, Johnson, and Albert); acquisition of data (Drs Johnson and Killiany and Mr Zimmerman); analysis and interpretation of data (Drs El Fakhri, Kijewski, and Albert); drafting of the manuscript (Drs El Fakhri, Kijewski, and Albert); critical revision of the manuscript for important intellectual content (Drs El Fakhri, Kijewski, Johnson, and Albert and Mr Zimmerman); statistical expertise (Drs El Fakhri and Kijewski); obtained funding (Drs Kijewski, Johnson, and Albert); administrative, technical, and material support (Dr El Fakhri and Messrs Syrkin, Becker, and Zimmerman); study supervision (Drs El Fakhri and Kijewski).

This work was supported in part by grants RO1-EB000802, PO1-AG04953, and RO1-CA78936 from the National Institutes of Health, Bethesda, Md.

We thank Mary Hyde, PhD, for data management and Kenneth Jones, EdD, for assistance with statistical analysis. The manuscript benefited greatly from the suggestions of 2 anonymous reviewers.

References
1.
Jagust  WJBudinger  TFReed  BR The diagnosis of dementia with single photon emission computed tomography. Arch Neurol.1987;44:258-262.
PubMed
2.
Johnson  KADavis  KRBuonanno  FSBrady  TJGrowdon  JH Comparison of magnetic resonance and x-ray computed tomography in dementia. Arch Neurol.1987;44:1075-1080.
PubMed
3.
DeKosky  SShih  WJScmitt  FCoupal  JKirkpatrick  C Assessing utility of single photon emission computed tomography (SPECT) scan in Alzheimer's disease: correlation with cognitive severity. Alzheimer Dis Assoc Disord.1990;4:14-23.
PubMed
4.
Holman  BLJohnson  KAGerada  BCarvalho  PASatlin  A The scintigraphic appearance of Alzheimer's disease: a prospective study using technetium-99m-HMPAO SPECT. J Nucl Med.1992;33:181-185.
PubMed
5.
Johnson  KJones  KHolman  BL  et al Preclinical prediction of Alzheimer's disease using SPECT. Neurology.1998;50:1563-1571.
PubMed
6.
Minoshima  SGiordani  BBerent  SFrey  KAFoster  NLKuhl  DE Metabolic reduction in the posterior cingulate cortex in very early Alzheimer's disease. Ann Neurol.1997;42:85-94.
PubMed
7.
Okamura  NShinkawa  MArai  H  et al Prediction of progression in patients with mild cognitive impairment using IMP-SPECT. Nippon Ronen Igakkai Zasshi.2000;37:974-978.
PubMed
8.
Du  ASchuff  NAmend  D  et al MRI of entorhinal cortex and hippocampus in mild cognitive impairment and Alzheimer's disease. J Neurol Neurosurg Psychiatry.2001;71:441-447.
PubMed
9.
Convit  Ade Leon  MJTarshish  C  et al Specific hippocampal volume reductions in individuals at risk for Alzheimer's disease. Neurobiol Aging.1997;18:131-138.
PubMed
10.
De Toledo-Morrell  LGoncharova  IDickerson  BWilson  RSBennett  DA From healthy aging to early Alzheimer's disease: in vivo detection of entorhinal cortex atrophy. Ann N Y Acad Sci.2000;911:240-253.
PubMed
11.
Killiany  RJGomez-Isla  TMoss  MB  et al Use of structural magnetic resonance imaging to predict who will get Alzheimer's disease. Ann Neurol.2000;47:430-439.
PubMed
12.
Kaye  JASwihart  THowieson  D  et al Volume loss of the hippocampus and temporal lobe in healthy elderly persons destined to develop dementia. Neurology.1997;48:1297-1304.
PubMed
13.
Xu  YJack Jr  CRO'Brien  PC  et al Usefulness of MRI measures of entorhinal cortex versus hippocampus in AD. Neurology.2000;54:1760-1767.
PubMed
14.
Albert  MSMoss  MBTanzi  RJones  K Preclinical prediction of AD using neuropsychological tests. J Int Neuropsychol Soc.2001;7:631-639.
PubMed
15.
Daly  EZaitchik  DCopeland  MSchmahmann  JGunther  JAlbert  MS Predicting conversion to Alzheimer disease using standardized clinical information. Arch Neurol.2000;57:675-680.
PubMed
16.
Hughes  WJBerg  LDanziger  WLCoben  LAMartin  RL A new clinical scale for the staging of dementia. Br J Psychiatry.1982;140:566-572.
PubMed
17.
Folstein  MFolstein  SMcHugh  P "Mini-Mental State": a practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res.1975;12:189-198.
PubMed
18.
McKhann  GDrachman  DFolstein  MKatzman  RPrice  DStadlan  EM Clinical diagnosis of Alzheimer's disease: report of the NINCDS-ADRDA Work Group under the auspices of Department of Health and Human Services Task Force on Alzheimer's disease. Neurology.1984;34:939-944.
PubMed
19.
National Institute on Aging, Reagan Institute Working Group on Diagnostic Criteria for the Neuropathological Assessment of Alzheimer's Disease Consensus recommendations for the postmortem diagnosis of Alzheimer's disease. Neurobiol Aging.1997;18(suppl):S1-S2.
PubMed
20.
Killiany  RJMoss  MBAlbert  MSSandor  TJolesz  F Temporal lobe regions on magnetic resonance imaging identify patients with early Alzheimer's disease. Arch Neurol.1993;50:949-954.
PubMed
21.
Genna  SSmith  AP The development of ASPECT, an annular single crystal brain camera for high efficiency SPECT. IEEE Trans Nucl Sci.1988;35:654-658.
22.
El Fakhri  GMoore  SCMaksud  PAurengo  AKijewski  MF Absolute activity quantitation in simultaneous I-123/Tc-99m brain SPECT. J Nucl Med.2001;42:300-308.
PubMed
23.
El Fakhri  GKijewski  MFMoore  SC Absolute activity quantitation from projections using an analytical approach: comparison with iterative methods in Tc-99m and I-123 brain SPECT. IEEE Trans Nucl Sci.2001;48:768-773.
24.
Moore  SCKijewski  MFMuller  SPRybicki  FZimmerman  RE Evaluation of scatter compensation methods by their effects on parameter estimation from SPECT projections. Med Phys.2001;28:278-287.
PubMed
25.
Shepp  LAVardi  Y Maximum likelihood reconstruction for emission tomography. IEEE Trans Med Imaging.1982;1:113-122.
26.
Hudson  HMLarkin  RS Accelerated image reconstruction using ordered subsets of projection data. IEEE Trans Med Imaging.1994;13:601-609.
27.
Pelizzari  CAChen  GTYSpelbring  DRWeichselbaum  RRChen  CT Accurate three-dimensional registration of CT, PET, and/or MR images of the brain. J Comput Assist Tomogr.1989;13:20-26.
PubMed
28.
Holman  BLZimmerman  REJohnson  KA  et al Computer-assisted superimposition of magnetic resonance and high-resolution technetium-99m-HMPAO and thallium-201 SPECT images of the brain. J Nucl Med.1991;32:1478-1484.
PubMed
29.
Rencher  AC Multivariate Statistical Inference and Applications.  New York, NY: John Wiley & Sons Inc; 1998.
30.
Metz  CE ROC methodology in radiologic imaging. Invest Radiol.1986;21:720-733.
PubMed
31.
El Fakhri  GBuvat  IBenali  HTodd-Pokropek  ADi Paola  R Relative impact of scatter, collimator response, attenuation, and finite spatial resolution corrections in cardiac SPECT. J Nucl Med.2000;41:1400-1408.
PubMed
32.
Albert  MSMoss  MB Neuropsychological approach to preclinical identification of Alzheimer's disease.  In: Squire  L, Schacter  D, eds. Neuropsychology of Memory. New York, NY: Guilford Publications; 2002:248-262.
×