Meta-analysis based SVM classification enables accurate detection of Alzheimer's disease across different clinical centers using FDG-PET and MRI☆
Introduction
Imaging biomarkers such as regional atrophy as measured by structural magnetic resonance imaging (MRI), glucose hypometabolism as measured by [F18]fluorodeoxyglucose positron emission tomography (FDG-PET) or brain β-amyloid load as measured by PIB- (Pittsburgh compound B) and florbetaben-PET have been reported to be useful in diagnosis and/or differential diagnosis of dementia (Sabri et al., 1999, Sabri et al., 2008, Hoffman et al., 2000, Rosen et al., 2002, Buckner et al., 2005, Diehl et al., 2004, Jeong et al., 2005, Diehl-Schmid et al., 2007, Edison et al., 2007, Fung and Stoeckel, 2007, Jack et al., 2008, Schroeter et al., 2007, Schroeter et al., 2008, Schroeter et al., 2009, Schroeter and Neumann, 2011, Davatzikos et al., 2008, Klöppel et al., 2007, Barthel et al., 2011). Therefore, it is now being suggested to incorporate such imaging markers into criteria for in vivo diagnosis of dementia (Dubois et al., 2007, Kipps et al., 2009).
Most previous biomarker studies focused on one specific biomarker or compared sensitivity and specificity of different single biomarkers (Fung and Stoeckel, 2007, Davatzikos et al., 2008, Habeck et al., 2008, Klöppel et al., 2007, Chaves et al., 2009, Habert et al., 2009, Horn et al., 2009, Ramirez et al., 2009a). Obviously, the combination of biomarkers potentially offers further improvements, and statistical methods such as multivariate pattern analyses using support vector machine (SVM) classifications not only enable automatic classification using one specific biomarker, but also provide a tool to combine two or more different biomarkers within the same classification model. For example, it has been shown that combining information from FDG-PET and MRI substantially improves detection (Hinrichs et al., 2009, Zhang et al., 2011) and differentiation of Alzheimer's disease dementia (AD) and frontotemporal lobar degeneration (Dukart et al., 2011). The volume-of-interest (VOI) approach used in Dukart et al. (2011), although less sensitive compared to whole-brain classification when using a single modality, was far superior to whole-brain classification when combined information from FDG-PET and MRI was used. VOIs used in this study were extracted from two comprehensive systematic and quantitative meta-analyses investigating both dementia syndromes in very large cohorts with anatomical likelihood estimates (ALE) (Schroeter et al., 2007, Schroeter et al., 2009). Accordingly, these VOIs represent the prototypical networks affected by these diseases and are not biased to a specific dataset. Furthermore, the preprocessing algorithm (Dukart et al., 2011) was designed to overcome difficulties which occur due to the use of different scanner types and different scanning sequences with different scaling and resolution. However, the generalizability of the results of that study (Dukart et al., 2011) was highly limited due to the low number of subjects and because data from only one center were used for SVM classification.
The goal of the present study was to further validate the approach proposed by Dukart et al. (2011) and to assess its generalizability to data from multicenter studies. To achieve this, we applied the identical preprocessing and classification algorithm to two different datasets. Classification accuracy results using FDG-PET and MRI data from the Clinic of Cognitive Neurology at the University of Leipzig were compared to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (www.adni-info.org). To avoid a classification bias towards the ADNI data (because substantially more control subjects and patients were available in this database than in the Leipzig cohort), we restricted the number of subjects and patients included from this database to make the numbers comparable with the Leipzig cohort. This is important because otherwise a combined classifier from both datasets would have mainly learned the distribution of subject and patient data in the ADNI cohort due to the substantially higher amount of subjects while practically ignoring the single center data. The ADNI database is a free access database containing, besides comprehensive neuropsychological and clinical evaluation, FDG-PET and MRI data of AD patients and healthy control subjects. We hypothesized that multicenter MRI and FDG-PET data obtained using different scanner types and sequences might be used to improve the accuracy of AD diagnosis in single clinical centers. Furthermore, we hypothesized that it would be possible to confirm our recent findings – namely improvement of individual dementia diagnosis with SVM classification of combined MRI and FDG-PET data and the aforementioned VOI approach based on meta-analyses.
Section snippets
Subjects
We analyzed FDG-PET and T1-weighted MRI data of 21 patients (Table 1) with an early stage of probable AD, 14 patients with an early stage of FTLD and 13 control subjects. Patients were recruited from the Clinic of Cognitive Neurology at the University Hospital Leipzig. Probable AD was diagnosed according to the clinical NINCDS-ADRDA criteria (McKhann et al., 1984). Diagnosis of FTLD was based on clinical criteria suggested by Neary et al. (1998). The control group included subjects who visited
Clinical characteristics
Clinical characteristics are illustrated in Table 1. The chi-square test for independent samples did not reveal any statistical differences in sex between the four groups [χ²(3)=5.76; p=0.12]. The ANOVAs revealed significant between-group differences in age, CDR, MMSE and education. AD subjects in the ADNI dataset had, as expected, significantly lower MMSE scores compared with the ADNI control group [t(54)=10.7; p<0.001]. There was no significant difference in the MMSE between AD patients in
Discussion
In our study we investigated classification accuracies for detection of AD using FDG-PET, MRI or combined information from both imaging modalities in two independent cohorts. Thereby, we used VOIs centered to coordinates reported in a comprehensive meta-analysis investigating AD (Schroeter et al., 2009) as features for SVM classification. Very high accuracies for discrimination between AD patients and control subjects were obtained in both cohorts using only FDG-PET or combined information from
Acknowledgment
Juergen Dukart, Henryk Barthel, Arno Villringer, Osama Sabri and Matthias L. Schroeter are supported by the Leipzig Research Center for Civilization Diseases (LIFE) at the University of Leipzig. LIFE is funded by the European Union, the European Regional Development Fund (ERFD) and the Free State of Saxony within the framework of the excellence initiative. Matthias L. Schroeter is also supported by the German consortium for frontotemporal lobar degeneration, funded by the German Federal
References (37)
- et al.
Effects of hardware heterogeneity on the performance of SVM Alzheimer's disease classifier
Neuroimage
(2011) A fast diffeomorphic image registration algorithm
Neuroimage
(2007)- et al.
Cerebral amyloid-β PET with florbetaben (18F) in patients with Alzheimer's disease and healthy controls: a multicentre phase 2 diagnostic study
Lancet Neurology
(2011) - et al.
SVM-based computer-aided diagnosis of the Alzheimer's disease using t-test NMSE feature selection with feature correlation weighting
Neuroscience Letters
(2009) - et al.
Individual patient diagnosis of AD and FTD via high-dimensional pattern classification of MRI
Neuroimage
(2008) - et al.
Cerebral metabolic patterns at early stages of frontotemporal dementia and semantic dementia. A PET study
Neurobiology of Aging
(2004) - et al.
Decline of cerebral glucose metabolism in frontotemporal dementia: a longitudinal 18F-FDG-PET-study
Neurobiology of Aging
(2007) - et al.
Research criteria for the diagnosis of Alzheimer's disease: revising the NINCDS-ADRDA criteria
Lancet Neurology
(2007) - et al.
Differential effects of global and cerebellar normalization on detection and differentiation of dementia in FDG-PET studies
Neuroimage
(2010) - et al.
Multidimensional classification of hippocampal shape features discriminates Alzheimer's disease and mild cognitive impairment from normal aging
Neuroimage
(2009)
Multivariate and univariate neuroimaging biomarkers of Alzheimer's disease
Neuroimage
Spatially augmented LPboosting for AD classification with evaluations on the ADNI dataset
Neuroimage
Differential automatic diagnosis between Alzheimer's disease and frontotemporal dementia based on perfusion SPECT images
Artificial Intelligence in Medicine
Towards a nosology for frontotemporal lobar degenerations-a meta-analysis involving 267 subjects
Neuroimage
Neural networks in frontotemporal dementia—a meta-analysis
Neurobiology of Aging
Neural correlates of Alzheimer's disease and mild cognitive impairment: a systematic and quantitative meta-analysis involving 1351 patients
Neuroimage
Multimodal classification of Alzheimer's disease and mild cognitive impairment
NeuroImage
Molecular, structural, and functional characterization of Alzheimer's disease: evidence for a relationship between default activity, amyloid, and memory
Journal of Neuroscience
Cited by (115)
Diagnosis of Alzheimer's disease via optimized lightweight convolution-attention and structural MRI
2024, Computers in Biology and MedicineDeep learning-based classification of healthy aging controls, mild cognitive impairment and Alzheimer's disease using fusion of MRI-PET imaging
2023, Biomedical Signal Processing and ControlCitation Excerpt :In addition, the spatial patterns of brain atrophy-AD index with CSF biomarkers were combined to predict conversion from MCI to AD using SVM [24]. Further, FDG-PET and GM density values of all the voxels of selected ROIs and the average of all the voxels of selected ROIs were classified AD using SVM classifier [25]. The combined regional GM volume, regional average PET intensity, and CSF biomarkers as features, and proposed a matrix-similarity-based loss function for better classification using SVM [26].
A review on Alzheimer's disease classification from normal controls and mild cognitive impairment using structural MR images
2023, Journal of Neuroscience Methods
- ☆
Data used in preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (www.loni.ucla.edu/ADNI). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. ADNI investigators include (complete listing available at 〈http://www.loni.ucla.edu/ADNI/Collaboration/ADNI_Manuscript_Citations.pdf〉).