Sage Journals: Discover world-class research

Abstract

Presently, visual and quantitative approaches for image-supported diagnosis of dementing disorders rely on regional intensity rather than on connectivity measurements. Here, we test metabolic connectivity for differentiation between Alzheimer’s disease and frontotemporal lobar degeneration. Positron emission tomography with 18F-fluorodeoxyglucose was conducted in 47 patients with mild Alzheimer’s disease, 52 patients with mild frontotemporal lobar degeneration, and 45 healthy elderly subjects. Sparse inverse covariance estimation and selection were used to identify patterns of metabolic, inter-subject covariance on the basis of 60 regional values. Relative to healthy subjects, significantly more pathological within-lobe connections were found in the parietal lobe of patients with Alzheimer’s disease, and in the frontal and temporal lobes of subjects with frontotemporal lobar degeneration. Relative to the frontotemporal lobar degeneration group, more pathological connections between the parietal and temporal lobe were found in the Alzheimer’s disease group. The obtained connectivity patterns differentiated between two patients groups with an overall accuracy of 83%. Linear discriminant analysis and univariate methods provided an accuracy of 74% and 69%, respectively. There are characteristic patterns of abnormal metabolic connectivity in mild Alzheimer’s disease and frontotemporal lobar degeneration. Such patterns can be utilized for single-subject analyses and might be more accurate in the differential diagnosis of dementing disorders than traditional intensity-based analyses.

Keywords

Alzheimer's disease FDG PET frontotemporal lobar degeneration positron emission tomography multivariate analysis

Introduction

There is increasing evidence that neurodegenerative dementias represent diseases of neural networks, with distinct patterns of disconnection being found in different dementing disorders.¹ Thus, damage to specific neural tracts was reported in Alzheimer’s disease (AD) and frontotemporal dementia (FTD),² two most common types of presenile dementia. Similarly, functional magnetic resonance imaging (fMRI) studies have pronounced divergent changes in resting state networks between AD and FTD.³ So far, however, the connectivity information has not been tested for purposes of differential diagnosis.

Positron emission tomography with 18F-fluorodeoxyglucose (FDG-PET) is a well-established tool for early and differential diagnosis of dementia.⁴ Thus, dementia due to AD and frontotemporal lobar degeneration (FTLD) could be differentiated by means of FDG-PET with a high accuracy.^5,6 In this context, both visual and quantitative analysis of PET images relies on identification of a characteristic spatial pattern of hypometabolism. These methods take into account intensity, or magnitude, of regional FDG-PET measurements, while information about interregional covariance is not considered.

In accord with growing interest in research on brain connectivity, recent studies have addressed the issue of covariance, or functional connectivity,⁷ in FDG-PET data. To distinguish from functional connectivity as quantified with fMRI, hereafter we use the term “metabolic connectivity.”⁸ Thus, plausible patterns of inter-subject metabolic covariance have been reported in healthy individuals.^9–11 More relevant from the clinical perspective, a few recent FDG-PET studies have consistently reported abnormalities in metabolic covariance in patients with mild cognitive impairment (MCI)^12–14 and dementia due to AD.^13,15 Yet, none of the studies above addressed individual subject analysis, a prerequisite for clinical applications.

Thus, the aim of the present work was to explore the value of metabolic connectivity in the differential diagnosis of mild AD and FTLD. To this end, we applied connectivity modeling in the form of sparse inverse covariance estimation (SICE) and selection (SICS).

Materials and methods

Subjects

We retrospectively analyzed FDG-PET images from 47 patients with mild AD, 52 patients with mild FTLD, and 45 healthy elderly subjects (HS). Their demographic characteristics are summarized in Table 1. The FTLD group included 35 patients with behavioral and 17 patients with language variants of FTLD. The AD group consisted of 30 patients with probable AD according to the NINCDS-ADRDA criteria¹⁶ and 17 patients with MCI due to AD.¹⁷ Evidence of AD pathology in these subjects was ensured by means of an abnormal PET scan with ¹¹C-Pittsburgh compound B. Mild AD was defined as an MMSE score of ≥21 points.¹⁸ FTLD was diagnosed according to the Neary criteria.¹⁹ Only subjects with an MMSE score of ≥22 were included.²⁰ The HS group consisted of elderly subjects without relevant psychiatric or neurological symptoms and no evidence for cognitive impairment. They were recruited mainly via advertisements in local newspapers.

Table 1.

Demographic characteristics.

	AD	FTLD	HS
Number	47	52	45
Male/female	28/19	33/19	17/28
Age (years)^a	69.0 ± 9.4	64.6 ± 8.8	62.0 ± 10.5
MMSE (score)^a	24.5 ± 2.1	25.5 ± 1.8	–

AD: Alzheimer’s disease; FTLD: frontotemporal lobar degeneration; HS: healthy elderly subjects.

Mean ± standard deviation.

This study was carried out in accordance with the latest version of the Declaration of Helsinki after the consent procedures had been approved by the local ethics committee of the medical faculty at the Technische Universität München. Written informed consent to the study was obtained from HS. The patients gave oral informed consent, since the FDG-PET examination was a part of a diagnostic work-up for suspected neurodegenerative disorder.

PET data acquisition and preprocessing

Details of image acquisition and pre-processing are reported elsewhere.²⁰ In brief, data were acquired under standard resting conditions using a Siemens ECAT EXACT HR + PET scanner (CTI, Knoxville, Tennessee, USA). A sequence of three frames with duration of 10, 5, and 5 min was acquired starting 30 min post injection. FDG-PET images were spatially normalized to the Montreal Neurological Institute space using Statistical Parametric Mapping software (http://www.fil.ion.ucl.ac.uk/spm/) implemented in Matlab 2013b (MathWorks Inc.). The brain was subdivided into 116 cortical volumes of interests (VOIs) using automated anatomical labeling (AAL) atlas (http://www.gin.cnrs.fr). All VOIs within the frontal, parietal, and temporal lobes, i.e. those typically affected in AD and FTLD were included (Table 2). Mean values of tracer uptake were extracted from each VOI separately for the right and left hemisphere. These values were then intensity scaled to uptake in the whole brain, defined as mean intensity value of all 116 VOIs.²¹

Table 2.

VOI labels according to the AAL atlas.

Frontal lobe		Parietal lobe		Temporal lobe
1	Frontal_Sup_L	31	Parietal_Sup_L	43	Temporal_Sup_L
2	Frontal_Sup_R	32	Parietal_Sup_R	44	Temporal_Sup_R
3	Frontal_Sup_Orb_L	33	Parietal_Inf_L	45	Temporal_Pole_Sup_L
4	Frontal_Sup_Orb_R	34	Parietal_Inf_R	46	Temporal_Pole_Sup_R
5	Frontal_Mid_L	35	SupraMarginal_L	47	Temporal_Mid_L
6	Frontal_Mid_R	36	SupraMarginal_R	48	Temporal_Mid_R
7	Frontal_Mid_Orb_L	37	Angular_L	49	Temporal_Pole_Mid_L
8	Frontal_Mid_Orb_R	38	Angular_R	50	Temporal_Pole_Mid_R
9	Frontal_Inf_Oper_L	39	Precuneus_L	51	Temporal_Inf_L 8301
10	Frontal_Inf_Oper_R	40	Precuneus_R	52	Temporal_Inf_R 8302
11	Frontal_Inf_Tri_L	41	Cingulum_Post_L	53	Fusiform_L
12	Frontal_Inf_Tri_R	42	Cingulum_Post_R	54	Fusiform_R
13	Frontal_Inf_Orb_L			55	Hippocampus_L
14	Frontal_Inf_Orb_R			56	Hippocampus_R
15	Olfactory_L			57	ParaHippocampal_L
16	Olfactory_R			58	ParaHippocampal_R
17	Frontal_Sup_Medial_L			59	Amygdala_L
18	Frontal_Sup_Medial_R			60	Amygdala_R
19	Frontal_Med_Orb_L
20	Frontal_Med_Orb_R
21	Rectus_L
22	Rectus_R
23	Cingulum_Ant_L
24	Cingulum_Ant_R
25	Rolandic_Oper_L
26	Rolandic_Oper_R
27	Supp_Motor_Area_L
28	Supp_Motor_Area_R
29	Cingulum_Mid_L
30	Cingulum_Mid_R

L: left, R: right hemisphere.

Metabolic connectivity pattern

We considered a dataset $X \in ℝ n \times p$ , where p variables denote regional tracer uptake values in selected VOIs and n samples are the number of acquired FDG-PET images. In analogy with previous studies,^11,22 the data were assumed to follow a multivariate Gaussian distribution. That is, $x_{1}, \dots, x_{n} ~ N (μ, Σ)$ , where $x_{i}$ is a p-dimensional vector, i = 1,…, n, $μ \in ℝ p$ is a mean vector, and $Σ \in ℝ p \times p$ is a true covariance matrix that is unknown and has to be estimated. Let $Θ = Σ - 1$ be a true inverse covariance (IC) matrix. Thereafter, we refer to an estimate $\overset{\land}{Θ}$ as a metabolic connectivity pattern. The standard approach to derive $\overset{\land}{Θ}$ is maximum likelihood estimation (MLE) of the true IC matrix $Θ$ . However, reliable MLE requires $n >> p$ . Otherwise, the estimation method is ill-posed, carrying a large estimation error.²³ Therefore, a sparse estimation of $Θ$ is preferable.^11,15 In brief, a sparse estimation of $Θ$ is robust when enough entries in $\overset{\land}{Θ}$ are set to zero. The amount of sparsity is controlled by regularization parameter λ. For the analyses below and visualization, we set a sparsity level of $\overset{\land}{Θ}$ equal to 0.1. That is, we explicitly set λ so that 10% of all entries in $\overset{\land}{Θ}$ are included in the analysis. For further classification, λ is derived using a maximum a posteriori probability (MAP) estimation method (see supplementary material for details). SICE is of particular advantage for studies with small n.^24,25 However, due to the effect of shrinkage SICE is insufficiently robust for estimating values of non-zero entries, providing only a good estimation of (sparse) binary structure of $\overset{\land}{Θ}$

struct (\hat{Θ})_{ij} = {1, if {\hat{θ}}_{ij} \neq 0 0, if {\hat{θ}}_{ij} = 0

To overcome this limitation, i.e. to estimate magnitudes of non-zero entries, we applied a constrained optimization algorithm²⁶ that fits the covariance to the predefined sparse binary structure of the estimated IC matrix. This procedure was denoted as SICS. SICS uses the principle of parsimony in parametric model fitting, and estimates covariance according to maximum entropy, maximum likelihood, existence and uniqueness. An example of the metabolic connectivity pattern $\overset{\land}{Θ}$ with estimated coefficients ${\overset{\land}{θ}}_{ij}$ for p = 7 regions is shown in Figure 1. Details about SICE and SICS are reported in the supplementary material.

Figure 1.

Metabolic connectivity pattern $\overset{\land}{Θ}$ for p = 7 regions. Numbers 1–7 that denote rows and columns in the pattern correspond to the one from the p = 7 regions. A non-zero element in the pattern is equal to estimated inverse covariance between connected regions. A zero element indicates connectivity that is not robust enough when sparse estimation by limited samples is performed.

Metabolic correlation network is modeled from the sparse binary structure of the metabolic connectivity pattern $struct (\overset{\land}{Θ})$ ; it can be represented as an undirected weighted graph. Graph nodes correspond to selected brain regions. Edge weights represent strengths of metabolic correlations between nodes as estimated by Pearson’s correlation coefficient (PCC) ${\overset{\land}{P}}_{ij}$ from $\overset{\land}{Σ} = \overset{\land}{Θ} - 1$

{\overset{\land}{P}}_{ij} = \frac{{\overset{\land}{σ}}_{ij}}{\sqrt{{\overset{\land}{σ}}_{ii} {\overset{\land}{σ}}_{jj}}}

where

{\overset{\land}{P}}_{ij}

are off-diagonal elements with i = 1,…, p, j = 1,…, p, and i ≠ j;

{\overset{\land}{σ}}_{ij}

are elements of matrix

\overset{\land}{Σ}

. The metabolic correlation network can be represented graphically in at least two ways (Figure 2).

Figure 2.

Metabolic correlation network: matrix representation (a), graph representation (b). Numbers 1–7 that denote rows and columns in the matrix representation and nodes in the graph representation correspond to the one from the p = 7 regions. Elements in the matrix representation and edges in the graph representation denote estimated correlation coefficients between connected regions.

Metabolic differentiation pattern

To characterize pathological changes in metabolic connectivity patterns and corresponding metabolic correlation networks of the AD and FTLD groups, we applied the algorithm presented in Figure 3. The null-hypothesis H₀ stated that there was no pairwise difference between correlations ${\overset{\land}{P}}_{ij}^{(1)}$ and ${\overset{\land}{P}}_{ij}^{(2)}$ at the same position (ij). Otherwise, the alternative hypothesis H₁ was accepted. After Fisher's z-transformation of PCCs, z-coefficients were compared between the patients and HS group. T-parameters of the Student’s distribution for each compared pair of coefficients were calculated. p < 0.05 after correction by means of false discovery rate (FDR)²⁷ was considered significant.

Figure 3.

Calculation of the metabolic differentiation pattern in the AD group; magenta and green cells denote pathological (i.e. significantly altered) and healthy connections (correlations), respectively.

In addition, we counted the number of pathologically altered connections within and between lobes. Permutation tests with replacement²⁸ were performed in each group. Finally, a two-sample t-test was applied for hypothesis testing. We accepted the null-hypothesis H₀ that the number of connections for the group 1 (AD) is equal to their number for the group 2 (FTLD): n₁ _ij = n₂ _ij , where i and j are the one of the lobes (frontal, parietal, or temporal). The alternative hypothesis H₁ was n₁ _ij ≠ n₂ _ij . Computed p-values were adjusted for multiple tests by FDR.

Single-subject analysis

Here, metabolic connectivity pattern $\overset{\land}{Θ}$ was used to differentiate between any two groups of subjects. After calculation of connectivity patterns ${\overset{\land}{Θ}}_{AD} and {\overset{\land}{Θ}}_{FTLD}$ as described above, the following algorithm was applied. Let a new subject with unclear diagnosis (AD vs. FTLD) have an FDG-PET scan $x = [x_{1}, \dots, x_{p}]$ , where x_i is tracer uptake in region i, i = 1,…, p, and p is the total number of investigated brain regions. Both the vector x and given datasets $X_{AD}$ and $X_{FTLD}$ follow a multivariate Gaussian distribution. The probability density of a new subject's scan x given a class $P (x | {\overset{\land}{Θ}}_{class})$ can be computed using the equation for a multivariate Gaussian distribution. Thus, the likelihood functions of x having AD and FTLD can be written as

P (x | {\overset{\land}{Θ}}_{AD}) = \frac{| {\overset{\land}{Θ}}_{AD} | \frac{1}{2}}{(2 π) \frac{p}{2}} \times exp (- \frac{(x - μ_{AD}) {\overset{\land}{Θ}}_{AD} (x - μ_{AD}) T}{2}),

P (x | {\overset{\land}{Θ}}_{FTLD}) = \frac{| {\overset{\land}{Θ}}_{FTLD} | 1 / 2}{(2 π) p / 2} \times exp (- \frac{(x - μ_{FTLD}) {\overset{\land}{Θ}}_{FTLD} (x - μ_{FTLD}) T}{2}),

where

μ_{AD}

and

μ_{FTLD}

are the mean vectors of the datasets

X_{AD}

and

X_{FTLD}

, respectively.

The new subject x is classified as AD, if $P (x | {\overset{\land}{Θ}}_{AD}) > P (x | {\overset{\land}{Θ}}_{FTLD})$ , and as FTLD otherwise. The same procedure was applied for differentiation of patients from healthy subjects. Of note, this procedure can be easily adopted for classification of three and more groups.

Reference methods

Diagnostic performance of the present methodology was compared with that of a widely used multivariate method such as linear discriminant analysis (LDA)^3,6,29 and with a straightforward univariate method based on the t-test.³⁰ In LDA, covariance was estimated using MLE. In the univariate method, the region with the largest difference between two groups was used for differentiation.³⁰ In addition, we performed SICE alone¹⁵ and a regression tree.^31–33 The former was conducted on the analogy of Huang et al.¹⁵ with a difference that the regularization parameter $λ$ was derived using MAP. In the regression tree, all regions those intensities significantly differed between the groups (t-test, p < 0.01, FDR corrected) were included in the classification. Like the simple univariate method above, this technique does not take into account the covariance information.

The same data were used as input to all statistical analyses. In all analyses leave-one-out cross-validation was applied.

Results

Metabolic differentiation patterns are presented in Figures 4 and 5. A pathological metabolic connection indicates that the correlation between two given regions significantly (p < 0.01 FDR corrected) differs from the same correlation in the HS group. In the AD group, pathologically altered were on average 6%, 60%, and 27% of connections within the frontal, parietal, and temporal lobe, respectively. In the FTLD group, corresponding values were 22%, 0%, and 52%. Distribution of abnormal connections within and between lobes is presented in Figure 6. Relative to the HS group, more pathological within-lobe connections were found in the parietal lobe of AD subjects, while in the frontal and temporal lobes of FTLD subjects (p < 0.01 FDR corrected). In addition, more pathological connections between the parietal and temporal lobe were found in the AD relative to FTLD group (p < 0.01 FDR corrected).

Figure 4.

Metabolic differentiation patterns for AD, FTLD, and HS groups in matrix representation with a sparsity level of 0.1, i.e. 10% of all entries are considered for sparse estimation of IC matrix. Each row (column) in the matrix corresponds to one of the 60 VOIs. A colored cell corresponds to an edge, i.e. to a metabolic connection between two VOIs. Each matrix contains internally three highlighted black squares that denote brain lobes. Magenta and green cells denote pathological (i.e. significantly altered) and healthy connections (correlations), respectively.

Figure 5.

Metabolic differentiation patterns for AD (a) and FTLD (b) groups with a sparsity level of 0.1 is overlaid onto a standard MRI brain template. Altered metabolic connections are represented by magenta lines, healthy connections – by green lines, and centers of VOIs – by black dots.

Figure 6.

Distribution of pathological metabolic connections as presented by box-plots. A star (*) denotes a significant difference in the number of pathological connections between the AD and FTLD groups (p < 0.01 FDR corrected).

Results of the differentiation by means of the proposed approach, LDA and univariate method are summarized in Table 3. As compared to the univariate method, LDA provided a higher accuracy for differentiation of AD vs. HS and vs. FTLD but not for HS vs. FTLD. Accuracy of differentiation was highest for the SICE/SICS in all comparisons, reaching 83% for AD vs. FTLD; corresponding values for LDA and univariate method were 74% and 69%, respectively. Performance of SICE alone and regression tree is reported in the supplementary material. Overall, SICE appeared less accurate than the combination of SICE and SICS but more accurate than all other methods. The regression tree provided nearly identical results to those of the simple univariate method.

Table 3.

Differentiation between the groups using three methods.

Differentiation	SICE + SICS	LDA	Univariate
AD vs. HS
Accuracy, %	95.7	88.0	83.7
Sensitivity, %	91.1	89.1	70.2
Specificity, %	100.0	87.0	97.8
FTLD vs. HS
Accuracy, %	85.6	74.2	84.5
Sensitivity, %	96.2	80.0	75.0
Specificity, %	73.4	69.2	95.5
AD vs. FTLD
Accuracy, %	82.8	73.7	68.7
Sensitivity, %	76.6	71.4	38.3
Specificity, %	88.4	76.0	96.1

AD: Alzheimer’s disease; FTLD: frontotemporal lobar degeneration; HS: healthy elderly subjects.

Discussion

Here, we report an optimized methodological approach for modeling metabolic connectivity at a group level, with a subsequent single-subject analysis. This covariance-based method was applied to differentiate between patients with mild AD and FTLD. Characteristic patterns of abnormal metabolic connectivity were found in both clinical entities. The proposed approach enabled a more accurate differentiation between two groups of patients than common uni- and multivariate methods.

So far, very few studies have examined metabolic connectivity/covariance in AD and none in FTLD. Morbelli et al.¹² reported reduced correlations of a cluster of AD-typical hypometabolism, prefrontal cortex, and hippocampus in probable AD. A voxel-wise seed-based method was adopted by Carbonell et al.¹⁴ They found aberrant metabolic connectivity in ß-amyloid-positive subjects with MCI. Similar to our study, Sanabria-Diaz et al.¹³ applied brain parcellation into a number of anatomical regions followed by construction of a metabolic connectivity pattern. The researchers revealed aberrant global and local network properties in patients with probable AD and MCI. Yet, the applied approaches do not allow individual subjects analyses as those required in clinical settings. We are aware of only one study that has tested the value of metabolic connectivity at a single-subject basis.¹⁵ Using SICE, the authors could differentiate patients with probable AD from healthy subjects with a reasonable accuracy. Their approach provided a binary matrix, thus being insensitive to continuous changes in metabolic connectivity. To estimate magnitudes of non-zero entries in the IC matrix, we applied SICS. This method has been successfully utilized in the field of fMRI,³⁴ but so far it has not been adopted for PET image analyses. Of note, SICS (following SICE) allowed estimation of differences in metabolic connectivity strength and improved accuracy of differentiation between the groups. As compared to Huang et al.¹⁵ we in addition applied an algorithm for optimal selection of the regularization parameter λ instead of its simple fixation. Not unexpectedly, higher (96% vs. 88%) accuracy of differentiation between patients with AD and healthy subjects was achieved in the present study, although more severely affected patients were included in Huang et al.¹⁵

Using the proposed approach, we found altered within-lobe connections to be distributed according to the known pattern of regional involvement. That is, pathological connections were most numerous in the parietal lobe in AD, while in the frontal and temporal lobes in FTLD. In a comparison between the patients groups, more altered within-lobe frontal and temporal connections were found in FTLD, in line with a more pronounced hypometabolism in this clinical entity.^35,36 Intriguing results were obtained by analyses of between-lobe connections. Whereas the number of abnormal frontoparietal and frontotemporal connections did not differ between AD and FTLD, significantly more abnormal parietotemporal connections were found in AD. In accord with this observation, integrity of the descending cingulate bundle, a major white matter tract connecting the medial parietal and temporal lobes,³⁷ was found to be more severely affected in AD than in FTLD, with the difference being non-significant though.²

As major focus of this study, we tested the proposed covariance-based sparse approach for purposes of differential diagnosis at a single subject level. The method achieved accuracy of 83% in discrimination between mild AD and FTLD. As expected, higher accuracy was found in discrimination of patients from healthy subjects, i.e. 96% for AD and 86% for FTLD. Overall, these results are comparable with previous FDG-PET studies.^5,6,38 Given a high heterogeneity in study design and methodology, a numerical comparison with the literature would be inappropriate. For instance, an important feature of our work is very mild disease severity. That is, in the studies reported so far, the range of the mean MMSE scores was 14.0–22.3 and 15.5–22.5 points for AD and FTLD populations, respectively (compare with 24.5 and 25.5 points in our study). Hence, we confronted these results with established quantitative methods such as univariate analysis and LDA. The proposed approach outperformed both methods. Indeed, as compared with the univariate analysis, the proposed algorithm considers not only information about regional intensity but mainly covariance information. In terms of SICE, the regional information is represented by the estimated vector of mean regional intensities, while covariance information by the estimated IC matrix. Whereas LDA can also be regarded as covariance method, it involves MLE. As in detail described in Materials and methods section and in supplementary material, reliable MLE requires the data sample size n to be substantially larger than the number of variables p, in our case VOIs, i.e. $n > \frac{1}{2} p (p + 1)$ . This condition is rarely satisfied in PET studies, where samples of n < 50 are commonly used. In contrast, even a rather coarse brain parcellation scheme such as AAL provides 116 VOIs. In this case MLE is ill-posed,³⁹ while sparse multivariate analysis adopted here robustly estimates covariance even for $n < p$ .²⁴

Visual image rating in clinical routine is supported by univariate techniques as implemented in popular image analysis tools like NEUROSTAT/3D-SSP (http://128.208.140.75/∼Download) and SPM (http://www.fil.ion.ucl.ac.uk/spm). Although the power of multivariate techniques in research settings is well established, their application in the daily routine has been problematic. So far, one group of studies has segmented whole brain glucose metabolism into implicit clusters using principal or independent component analysis, followed by MLE of covariance.^40–42 The analysis of only a few main components allows avoiding large errors in covariance estimation when applied to limited PET samples. However, the resultant clusters do typically not correspond to anatomical regions, limiting interpretation and consequently, clinical application. In another group of studies like in our one, the brain is partitioned into explicit regions, i.e. anatomical structures according to various brain atlases.^8,43 Here, the number of employed VOIs is usually larger than the number of available PET images, precluding accurate MLE of covariance. Therefore, these studies are limited to (explorative) correlation analyses. In contrast and as mentioned above, the method applied in the current study is robust to small sample sizes. In addition, the approach generates a metabolic differentiation pattern in the form of a color map that is based on proper anatomical regions (Figure 4). These advantages favor testing this method in clinical settings.

As a major study limitation, we defined pathological connections as correlations whose strength was significantly different from that in the healthy state. Such definition corresponds to the nature of the connectivity pattern, which captures the relationships as a whole, i.e. pattern, and does not discriminate between stronger/weaker covariance/correlations. Basically, such discrimination is possible, but more subjects (at least n > p) would be necessary for reliable estimation of the IC matrix with more elements. Future studies with larger datasets should explore the direction of changes in correlation strength. Further, to minimize inflated accuracy due to data overfitting, the robustness of classification was tested using a leave-one-out cross-validation. Although this technique is recommended for evaluating classifier performance with small samples, an independent dataset would allow a more robust validation. Yet, this limitation concerns to the same or similar extent all methods, such that the comparison between the methods should actually not be biased. Further, a detailed neuropsychological testing was not available for most HS. Consequently, we cannot exclude that some HS had very mild cognitive deficits that were not evident clinically and not perceived by subjects themselves. Thus, the provided specificity values might be underestimated. Finally, gender distribution differed between the patients (about 3/5 males) and healthy (about 2/5 males) groups. Hence, putative, gender-associated variations in brain glucose metabolism might have biased the results. Again, however, two last reservations apply to all statistical methods tested, not favoring the proposed approach.

Footnotes

Funding

The study was supported by the internal grant program for resident physicians (KKF, project number B23-13/8764179) at the Technische Universität München (IY).

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Authors’ contributions

Design or conceptualization of the study: DT, KS, IY; Analysis or interpretation of the data: DT, KS, NZ, JL, IY; Drafting or revising the manuscript for intellectual content: all authors.

Supplementary material

Supplementary material for this paper can be found at

References

Pievani

Filippini

van den Heuvel

. Brain connectivity in neurodegenerative diseases – from phenotype to proteinopathy. Nat Rev Neurol 2014; 10: 620–633.

Zhang

Schuff

. White matter damage in frontotemporal dementia and Alzheimer's disease measured by diffusion MRI. Brain 2009; 132(Pt 9): 2579–2592.

Zhou

Greicius

Gennatas

. Divergent network connectivity changes in behavioural variant frontotemporal dementia and Alzheimer's disease. Brain 2010; 133(Pt 5): 1352–1367.

Bohnen

Djang

DSW

Herholz

. Effectiveness and safety of 18F-FDG PET in the evaluation of dementia: a review of the recent literature. J Nucl Med 2012; 53: 59–71.

Forster

Heidebrink

Clark

. FDG-PET improves accuracy in distinguishing frontotemporal dementia and Alzheimer’s disease. Brain 2007; 130: 2616–2635.

Mosconi

Tsui

Herholz

. Multicenter standardized 18F-FDG PET diagnosis of mild cognitive impairment, Alzheimer's disease, and other dementias. J Nucl Med 2008; 49: 390–398.

Friston

. Functional and effective connectivity in neuroimaging: a synthesis. Hum Brain Mapp 1994; 1–2: 56–78.

Lee

Kang

Kim

. Metabolic connectivity by interregional correlation analysis using statistical parametric mapping (SPM) and FDG brain PET; methodological development and patterns of metabolic connectivity in adults. Eur J Nucl Med Mol Imaging 2008; 35: 1681–1691.

Biswal

. Alzheimer's disease neuroimaging initiative. Metabolic brain covariant networks as revealed by FDG-PET with reference to resting-state fMRI networks. Brain Connect 2012; 2: 275–283.

10.

Yakushev

Chételat

Fischer

. Metabolic and structural connectivity within the default mode network relates to working memory performance in young healthy adults. Neuroimage 2013; 79: 184–190.

11.

Zou

Chetelat

Baydogan

. Metabolic connectivity as index of verbal working memory. J Cereb Blood Flow Metab 2015; 35: 1122–1126.

12.

Morbelli

Drzezga

Perneczky

. Resting metabolic connectivity in prodromal Alzheimer's disease. A European Alzheimer Disease Consortium (EADC) project. Neurobiol Aging 2012; 33: 2533–2550.

13.

Sanabria-Diaz

Martínez-Montes

Melie-Garcia

. Alzheimer’s Disease Neuroimaging Initiative. Glucose metabolism during resting state reveals abnormal brain networks organization in the Alzheimer's disease and mild cognitive impairment. PloS one 2013; 8: e68860.

14.

Carbonell

Charil

Zijdenbos

. β-Amyloid is associated with aberrant metabolic connectivity in subjects with mild cognitive impairment. J Cereb Blood Flow Metab 2014; 34: 1169–1179.

15.

Huang

Sun

. Learning brain connectivity of Alzheimer's disease by sparse inverse covariance estimation. Neuroimage 2010; 50: 935–949.

16.

McKhann

Drachman

Folstein

. 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.

17.

Albert

DeKosky

Dickson

. The diagnosis of mild cognitive impairment due to Alzheimer's disease: recommendations from the National Institute on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease. Alzheimer's Dementia 2011; 7: 270–279.

18.

Shin

Carter

Masterman

. Neuropsychiatric symptoms and quality of life in Alzheimer disease. Am J Geriatr Psychiatry 2005; 13: 469–474.

19.

Neary

Snowden

Gustafson

. Frontotemporal lobar degeneration: a consensus on clinical diagnostic criteria. Neurology 1998; 51: 1546–1554.

20.

Dukart

Perneczky

Förster

. Reference cluster normalization improves detection of frontotemporal lobar degeneration by means of FDG-PET. PloS One 2013; 8: e55415.

21.

Spetsieris

Eidelberg

. Scaled subprofile modeling of resting state imaging data in Parkinson's disease: methodological issues. Neuroimage 2011; 54: 2899–2914.

22.

Huang

. Brain Effective Connectivity Modeling for Alzheimer’s Disease by Sparse Gaussian Bayesian Network. KDD August San Diego 2011; 11: 21–24.

23.

Ledoit

Wolf

. A well-conditioned estimator for large-dimensional covariance matrices. J Multivariate Anal 2004; 88: 365–411.

24.

Friedman

Hastie

Tibshirani

. Sparse inverse covariance estimation with the graphical lasso. Biostatistics 2008; 9: 432–441.

25.

Yuan

Lin

. Model selection and estimation in the Gaussian graphical model. Biometrika 2007; 94: 19–35.

26.

Dempster

. Covariance selection. Biometrics 1972; 28: 157–175.

27.

Yekutieli

Benjamini

. Resampling-based false discovery rate controlling multiple test procedures for correlated test statistics. Plan Infer 1999; 82: 171–196.

28.

Good

. Permutation, parametric, and bootstrap tests of hypotheses, 3rd edn. New York: Springer, 2005.

29.

Yakushev

Landvogt

Buchholz

. Choice of reference area in studies of Alzheimer's disease using positron emission tomography with fluorodeoxyglucose-F18. Psychiatry Res 2008; 164: 143–153.

30.

Habeck

Stern

. Alzheimer’s disease neuroimaging initiative. Multivariate data analysis for neuroimaging data: overview and application to Alzheimer's disease. Cell Biochem Biophys 2010; 58: 53–67.

31.

Kodewitz

Lelandais

Montagne

. Alzheimer's disease early detection from sparse data using brain importance maps. Electronic Lett Comput Vision Image Anal 2013; 12: 42–56.

32.

Sadeghi N, Foster NL, Wang AY, et al. Automatic classification of Alzheimer’s disease vs. frontotemporal dementia: a spatial decision tree approach with FDG-PET. In: 5th IEEE international symposium on biomedical imaging: from nano to macro, Paris, France, 14–17 May 2008, vol. 1–4, pp. 408–411.

33.

Ramírez

Górriz

Segovia

. Computer aided diagnosis system for the Alzheimer's disease based on partial least squares and random forest SPECT image classification. Neurosci Lett 2010; 472: 99–103.

34.

Varoquaux

Poline

. A novel sparse graphical approach for multimodal brain connectivity inference. Med Image Comput Comput Assist Interv Int Conf Med Image Comput Comput Assist Interv 2012; 15(Pt 1): 707–714.

35.

Ishii

Sakamoto

Sasaki

. Cerebral glucose metabolism in patients with frontotemporal dementia. J Nucl Med 1998; 39: 1875–1878.

36.

Santens

De Bleecker

Goethals

. Differential regional cerebral uptake of (18)F-fluoro-2-deoxy-D-glucose in Alzheimer's disease and frontotemporal dementia at initial diagnosis. Eur Neurol 2001; 45: 19–27.

37.

Catani

Michel Thiebaut de

Schotten

. Atlas of human brain connections, Oxford: Oxford University Press, 2012.

38.

Rabinovici

Rosen

Alkalay

. Amyloid vs FDG-PET in the differential diagnosis of AD and FTLD. Neurology 2011; 77: 2034–2042.

39.

Ledoit

Wolf

. A well-conditioned estimator for large-dimensional covariance matrices. J Multivariate Anal 2004; 88: 365–411.

40.

Habeck

Foster

Perneczky

. Multivariate and univariate neuroimaging biomarkers of Alzheimer's disease. Neuroimage 2008; 40: 1503–1515.

41.

Salmon

Kerrouche

Herholz

. Decomposition of metabolic brain clusters in the frontal variant of frontotemporal dementia. Neuroimage 2006; 30: 871–878.

42.

Toussaint

Perlbarg

Bellec

. Resting state FDG-PET functional connectivity as an early biomarker of Alzheimer's disease using conjoint univariate and independent component analyses. Neuroimage 2012; 63: 936–946.

43.

Azari

Rapoport

Grady

. Patterns of interregional correlations of cerebral glucose metabolic rates in patients with dementia of the Alzheimer type. Neurodegeneration 1992; 1: 101–111.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.04 MB

Metabolic connectivity for differential diagnosis of dementing disorders

Abstract

Keywords

Introduction

Materials and methods

Subjects

PET data acquisition and preprocessing

Metabolic connectivity pattern

Metabolic differentiation pattern

Single-subject analysis

Reference methods

Results

Discussion

Footnotes

Funding

Declaration of conflicting interests

Authors’ contributions

Supplementary material

References

Supplementary Material