C H E R U B I N I
28.
E T
A L
Schnoebelen S, Semrud-Clikeman M, Pliszka SR. Corpus callosum anatomy in chronically treated and stimulant na€ıve ADHD. J Atten Disord 2010;14:256-266.
Magnetic Resonance Support Vector Machine Discriminates Essential Tremor With Rest Tremor From Tremor-Dominant Parkinson Disease , MD,1† Andrea Cherubini, PhD,1† Rita Nistico Fabiana Novellino, MD,1 Maria Salsone, MD,2 Salvatore Nigro, PhD,1 Giulia Donzuso, MD1 and Aldo Quattrone, MD1,2* 1
Neuroimaging Research Unit, Institute of Molecular Bioimaging and Physiology-National Research Council, Catanzaro, Italy and; 2 Institute of Neurology, Magna Graecia University, Catanzaro, Italy
ABSTRACT Background: The aim of the current study was to distinguish patients who had tremor-dominant Parkinson’s disease (tPD) from those who had essential tremor with rest tremor (rET). Methods: We combined voxel-based morphometryderived gray matter and white matter volumes and diffusion tensor imaging-derived mean diffusivity and fractional anisotropy in a support vector machine (SVM) to evaluate 15 patients with rET and 15 patients with tPD. Dopamine transporter single-photon emission computed tomography imaging was used as ground truth. Results: SVM classification of individual patients showed that no single predictor was able to fully discriminate patients with tPD from those with rET. By contrast, when all predictors were combined in a multimodal algorithm, SVM distinguished patients with rET from those with tPD with an accuracy of 100%. Conclusions: SVM is an operator-independent and automatic technique that may help distinguish patients with C 2014 tPD from those with rET at the individual level. V International Parkinson and Movement Disorder Society Key Words: resting tremor; magnetic resonance imaging; support vector machine; computer-aided diagnosis
therapeutic strategies and have different prognoses. In this context, dopamine transporter single-photon emission computed tomography (DAT-SPECT) plays a major role, yet it represents a highly invasive and costly diagnostic procedure.3 Over the past decade, advanced magnetic resonance imaging (MRI) methodologies have been applied to group-level image analyses with the aim of quantifying the pattern of structural alterations between patients with PD, patients with ET, and healthy controls.4-6 Those studies have shown brain atrophy in many cortical and subcortical regions, particularly in the basal ganglia, in the substantia nigra of patients with PD, and in the cerebellum of patients with ET. However, currently, no data exists on MRI in patients with rET, and only a few studies have focused on MRI in patients with tPD. Some authors observed an increased gray matter (GM) volume in the nucleus ventralis intermedius of the thalamus in patients with tPD compared with healthy controls,7 whereas others reported reduced brain activity in the prefrontal cortex and globus pallidus of patients who had non-tPD compared with both patients who had tPD and healthy controls.8 Recently, support vector machines (SVMs) have been applied with excellent results to the analysis of MRI data for the computer-aided diagnosis of patients with different neurological diseases.9-16 SVMs differ from conventional MRI studies performed at a group level, because they can label each patient using a combination of multiple features extracted from individual images. In the current study, we investigated the potential usefulness of SVMs for differentiating patients with tPD from those with rET using a combined whole-brain, voxel-based morphometry (VBM) and diffusion tensor imaging (DTI) analyses.
Patients and Methods We enrolled 15 patients who had a diagnosis of tPD according to established clinical criteria17 and 15 patients who had rET.18 In patients with tPD and rET, clinical evaluations and tremor scores were calculated according to the Unified Parkinson’s Disease Rating Scale motor examination (UPDRS-ME) (section III, items 20 and 21) and the Fahn-Tolosa scale. No
-----------------------------------------------------------*Correspondence to: Prof. Aldo Quattrone, Institute of Neurology, Magna Graecia University, Viale Europa, 88100, Catanzaro, Italy; [email protected] †
Differentiating between patients who have essential tremor (ET) with resting tremor (rET) from those who have tremor-dominant Parkinson’s disease (tPD) may be challenging in the early phases of the diseases.1,2 Distinguishing between these pathologies, however, is an important task, because they demand different
1216
Movement Disorders, Vol. 29, No. 9, 2014
The first two authors contributed equally to this work.
Relevant conflicts of interest/financial disclosures: Nothing to report. Full financial disclosures and author roles may be found in the online version of this article. Received: 31 July 2013; Revised: 31 January 2014; Accepted: 17 February 2014 Published online 13 April 2014 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/mds.25869
M R - S V M
patient had any history of thyroid disease, cerebrovascular disease, or other neurologic diseases. All participants underwent the same imaging protocol, including MRI and DAT-SPECT. Patients were examined using a 3-Tesla GE MR750 scanner (GE Healthcare, Rahway, NJ). The MRI protocol included whole-brain, three-dimensional, T1-weighted, spoiled gradient recall echo (voxel size 5 1 3 1 3 1 mm3); DTI (b 5 0,1000; 27 directions; voxel size 5 2 3 2 3 2 mm3); and conventional clinical sequences. Image processing was performed using the Functional Magnetic Resonance Imaging of the Brain (FMRIB) Analysis Group Software Library (FSL) (Oxford University, United Kingdom).19 T1-weighted volumes were analyzed using the FSL voxel-based morphometry (FSLVBM) tool, and DTI volumes were analyzed with FSLDTIfit. To compare structural patterns across individuals, each individual image was transformed into a common coordinate system, called standard stereotactic space. The modulated VBM-style segmented images (GM volume and white matter [WM] volume), the mean diffusivity (MD), and fractional anisotropy (FA) maps, all in standard space, represented the four features for each voxel that were fed to the SVM algorithm in order to classify single patients. Images in standard space were smoothed with a Gaussian kernel (6 mm) before statistical processing. Voxel-wise, conventional, univariate statistical analysis on each of the four predictors was performed in RANDOMISE, part of FSL. Age and gender were used as non-explanatory coregressors. The individual-level SVM analysis was performed with LIBSVM20 using the method described by Zhang et al.13 and Mwangi et al.14 for multimodal image markers. Although the SVM is capable of working with a large number of features—in our case, four features per voxel—not all of these will be equally relevant for the classification. Consequently, we preselected the most important brain regions by using feature selection. Machine learning studies have demonstrated that feature selection is important for achieving a high predictive accuracy classifier.21 We performed an F test during training to identify the voxels that differed mostly in rET versus tPD. An optimal threshold of P was identified by empirically varying the threshold using a cross-validation procedure with the training data.14 A multi-kernel approach was employed when all predictors were combined.13 Sensitivity and specificity were assessed by a leaveone-out strategy: each patient, in turn, was removed both from the training set and from the generation of the weighting map. The left-out patient was then classified using the trained model to evaluate the accuracy of the algorithm. Before inclusion in the study, written informed consent was obtained from all participants, and the institutional review board approved the study.
D I S C R I M I N A T E S
B E T W E E N
r E T
A N D
t P D
TABLE 1. Demographic, clinical, and imaging features in patients with tremor-dominant Parkinson’s disease and patients with essential tremor with rest tremor Characteristic
tPD, n 5 15
rET, n 5 15
P
Age: Mean 6 SD, y Men: No. (%) Age at onset: Mean 6 SD, y Duration of disease: Mean 6 SD, y Familial history: No. (%) UPDRS-ME score: Mean 6 SD Fahn-Tolosa score: Mean 6 SD DAT-SPECT:d Mean 6 SD Left putamen Right putamen
68.7 6 5.8 9 (60%) 61.2 6 6.7
68.1 6 8.8 8 (53) 58.3 6 8.7
NSa NSb NSa
7.7 6 3.0
9.9 6 7.1
NSa
4 (22%) 13.0 6 1.9
7 (46) 11.4 6 5.7
NSb NSc
18.9 6 5.4
18.6 6 11.8
NSc
1.04 6 0.16 1.08 6 0.14
2.07 6 0.30 2.15 6 0.30
< 0.001a < 0.001a
a
This P value was value determined using an unpaired t test. This P value was value determined using a v2 test. This P value was value determined using the Mann-Whitney test. Abbreviations: tPD, tremor-dominant Parkinson’s disease; rET, essential tremor with rest tremor; SD, standard deviation; NS, nonsignificant; UPDRS-ME, Unified Parkinson’s Disease Rating Scale-Motor Examination (items 20 and 21); DAT-SPECT, dopamine transporter single-photon emission computed tomography. d The DAT-SPECT value was determined using the ratio of specific to nonspecific radioligand binding. b c
Results Clinical, demographic, and DAT-SPECT characteristics of the tPD and rET groups are shown in Table 1. Patients with rET had normal DAT-SPECT imaging, whereas tracer uptake in patients with tPD was decreased in the putamina. VBM and DTI results from the conventional univariate F test are shown in Figure 1A at a statistical threshold of P < 0.001. No voxel survived correction for multiple comparisons in any of the four parameters considered (GM, WM, MD, and FA). Group differences at the uncorrected threshold between rET and tPD patients were localized in caudate nucleus, globus pallidus, midbrain, internal capsule, body of the corpus callosum, and cerebellum regions. The SVM algorithm showed a high accuracy for distinguishing rET from tPD when GM or WM was used as a single predictor, whereas FA and MD were less accurate. When the four predictors (GM, WM, FA, and MD) were used simultaneously in a multimodal SVM, the algorithm had an accuracy of 100% for distinguishing patients with rET from those with tPD, thus allowing a differential diagnosis at the individual level. A graphic representation of this analysis is provided in Figure 1B.
Discussion In the current study, for the first time, we used a combination of VBM and DTI analyses for the differentiation of patients with rET from those with tPD. Multi-modal pattern recognition of MRI data was
Movement Disorders, Vol. 29, No. 9, 2014
1217
C H E R U B I N I
E T
A L
FIG. 1. The results from a conventional univariate F test between patients who had tremor-dominant Parkinson’s disease (tPD) and patients who had essential tremor with resting tremor (rET) are superimposed on a standard template. (A) Selected axial slices show the significant (P < 0.001, uncorrected) voxels for different predictors, represented by the voxel-based morphometry (VBM)-derived gray matter (GM) and white matter (WM) volume and the diffusion tensor imaging (DTI)-derived mean diffusivity (MD) and fractional anisotropy (FA). (B) Performances of a support vector machine (SVM) classifier were obtained using the four voxel-based predictors. Color is indicative of the diagnosis according to clinical criteria and dopamine transporter single-photon emission computed tomography imaging: red indicates patients with rET, and cyan indicates patients with tPD. Patients are distributed along a vertical axis according to the probability of belonging to one of the two groups resulting from SVM classification. The far right column shows the performance of the classifier using all four predictors combined (All).
capable of fully discriminating between patients with tPD and patients with rET at the individual level. Although a differential diagnosis between rET and tPD is possible using DAT-SPECT imaging, this technique is invasive and expensive as a routine diagnostic tool. To our knowledge, there are currently no studies investigating the potential of SVM for the individual detection of rET, and only few studies have investigated computer-aided diagnosis of PD using MRI.10,12,15,16 Conventional analyses of MRI data have been conducted in patients with ET by studying VBM,22,23 DTI,6,24,25 and iron accumulation26 compared with healthy controls. Although most studies suggested an impairment of the cerebello-thalamocortical loop, the results remain controversial. Regarding rET, currently, there are no MRI studies focused on this clinical subtype of ET. Similarly, studies investigating brain structural alterations in patients with PD have not yet drawn any congruent conclusion. Some studies showed that there was no structural difference between patients with PD and controls, and others showed a nigrostriatal signature of PD and an impairment of the basal gangliathalamocortical loop.4,27,28
1218
Movement Disorders, Vol. 29, No. 9, 2014
Our results of conventional MRI univariate analyses are in agreement with the literature, although the limited sample size of our study did not allow for high statistical significance. Thus, we attempted to increase the reliability of MRI by combining VBM and DTI in an SVM algorithm with the aim of increasing the overall sensitivity of the analysis. Our results showed that the performances of single classifiers (GM, WM, FA, MD) were lower than the multi-modal method classifier (Fig. 1B). Indeed, when the four image predictors were fed simultaneously to the SVM algorithm, we observed an overall accuracy of 100% in distinguishing patients with tPD from those with rET. Several works have highlighted the finding that multimodal MRI is better at capturing the chain of structural alterations acting simultaneously on brain tissue at the macroscopic and microscopic levels.29,30 Indeed, this might be the reason why the best accuracy was observed only when all predictors were used by the SVM at the same time. There are some limitations to this study. First, this was a preliminary study with a small sample size, and the reported accuracy might vary when the method is employed on a larger population with more
M R - S V M
heterogeneous phenotypes. Second, we used the same patients in the training and testing of the SVM algorithm, a procedure that could lead to data over fitting and, consequently, to an overestimation of classification accuracy. To limit this problem, we applied a leaveone-out protocol, ensuring that the classifier was always tested on new data sets, ie, data sets that had not been used previously to build the classification model. In conclusion, SVM analysis of voxel-based MRI is an operator-independent and automatic postprocessing technique that may help to correctly classify patients with resting tremor at the individual level and in the absence of a DAT scan investigation. SVM may be used for the differentiation of patients who have clinically indistinguishable phenotypes, such as those with tPD and rET, thus representing a valid help in everyday clinical practice.
References 1.
Thenganatt MA, Louis ED. Distinguishing essential tremor from Parkinson’s disease: bedside tests and laboratory evaluations. Expert Rev Neurother 2012;12:687-696.
2.
Cohen O, Pullman S, Jurewicz E, Watner D, Louis ED. Rest tremor in patients with essential tremor: prevalence, clinical correlates, and electrophysiologic characteristics. Arch Neurol 2003;60: 405-410.
D I S C R I M I N A T E S
B E T W E E N
r E T
A N D
t P D
cation of mild cognitive impairment [serial online]. PLoS One 2012;7:e32441. 12.
Long D, Wang J, Xuan M, et al. Automatic classification of early Parkinson’s disease with multi-modal MR imaging [serial online]. PLoS One 2012;7:e47714.
13.
Zhang D, Wang Y, Zhou L, Yuan H, Shen D. Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage 2011;55:856-867.
14.
Mwangi B, Ebmeier KP, Matthews K, Steele JD. Multi-centre diagnostic classification of individual structural neuroimaging scans from patients with major depressive disorder. Brain 2012;135: 1508-1521.
15.
Haller S, Badoud S, Nguyen D, Garibotto V, Lovblad KO, Burkhard PR. Individual detection of patients with Parkinson disease using support vector machine analysis of diffusion tensor imaging data: initial results. AJNR Am J Neuroradiol 2012;33: 2123-2128.
16.
Haller S, Badoud S, Nguyen D, Barnaure I, Montandon ML, Lovblad KO, Burkhard PR. Differentiation between Parkinson disease and other forms of Parkinsonism using support vector machine analysis of susceptibility-weighted imaging (SWI): initial results. Eur Radiol 2013;23:12-19.
17.
Gelb DJ, Oliver E, Gilman S. Diagnostic criteria for Parkinson disease. Arch Neurol 1999;56:33-39.
18.
Deuschl G, Bain P, Brin M. Consensus statement of the Movement Disorder Society on tremor.Ad Hoc Scientific Committee. Mov Disord 1998;13(suppl 3):2-23.
19.
Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. Neuroimage 2012;62:782-790.
20.
Chang C-C, Lin C-J. LIBSVM: a library for support vector machines. ACM T Intel Syst Technol 2011;2:3-27.
21.
Guyon I, Elisseeff A. An introduction to variable and feature selection. J Mach Learn Res 2003;3:1157-1182.
3.
Badiavas K, Molyvda E, Iakovou I, Tsolaki M, Psarrakos K, Karatzas N. SPECT imaging evaluation in movement disorders: far beyond visual assessment. Eur J Nucl Med Mol Imaging 2011;38: 764-773.
22.
Benito-Leon J, Alvarez-Linera J, Hernandez-Tamames JA, AlonsoNavarro H, Jimenez-Jimenez FJ, Louis ED. Brain structural changes in essential tremor: voxel-based morphometry at 3-Tesla. J Neurol Sci 2009;287(1–2):138-142.
4.
Stoessl AJ, Martin WW, McKeown MJ, Sossi V. Advances in imaging in Parkinson’s disease. Lancet Neurol 2011;10:987-1001.
23.
5.
Passamonti L, Cerasa A, Quattrone A. Neuroimaging of essential tremor: what is the evidence for cerebellar involvement [serial online]? Tremor Other Hyperkinet Mov (NY) 2012;2:02-67-421-3.
Quattrone A, Cerasa A, Messina D, et al. Essential head tremor is associated with cerebellar vermis atrophy: a volumetric and voxelbased morphometry MR imaging study. AJNR Am J Neuroradiol 2008;29:1692-1697.
24.
6.
Prodoehl J, Li H, Planetta PJ, et al. Diffusion tensor imaging of Parkinson’s disease, atypical parkinsonism, and essential tremor. Mov Disord 2013;28:1816-1822.
Nicoletti G, Manners D, Novellino F, et al. Diffusion tensor MRI changes in cerebellar structures of patients with familial essential tremor. Neurology 2010;74:988-994.
25.
7.
Kassubek J, Juengling FD, Hellwig B, Spreer J, Lucking CH. Thalamic gray matter changes in unilateral Parkinsonian resting tremor: a voxel-based morphometric analysis of 3-dimensional magnetic resonance imaging. Neurosci Lett 2002;323:29-32.
Saini J, Bagepally BS, Bhatt MD, et al. Diffusion tensor imaging: tract based spatial statistics study in essential tremor. Parkinsonism Relat Disord 2012;18:477-482.
26.
Novellino F, Cherubini A, Chiriaco C, et al. Brain iron deposition in essential tremor: a quantitative 3-Tesla magnetic resonance imaging study. Mov Disord 2013;28:196-200.
27.
Peran P, Cherubini A, Assogna F, et al. Magnetic resonance imaging markers of Parkinson’s disease nigrostriatal signature. Brain 2010;133:3423-3433.
8.
Prodoehl J, Planetta PJ, Kurani AS, Comella CL, Corcos DM, Vaillancourt DE. Differences in brain activation between tremorand nontremor-dominant Parkinson disease. JAMA Neurol 2013; 70:100-106.
9.
Kloppel S, Abdulkadir A, Jack CR Jr, Koutsouleris N, MouraoMiranda J, Vemuri P. Diagnostic neuroimaging across diseases. Neuroimage 2012;61:457-463.
28.
Pan PL, Song W, Shang HF. Voxel-wise meta-analysis of gray matter abnormalities in idiopathic Parkinson’s disease. Eur J Neurol 2012;19:199-206.
10.
Cherubini A, Morelli M, Nistico R, et al. Magnetic resonance support vector machine discriminates between Parkinson disease and progressive supranuclear palsy [published online ahead of print 3 Dec 2013]. Mov Disord 2013. doi: 10.1002/mds.25737.
29.
Cherubini A, Peran P, Caltagirone C, et al. Aging of subcortical nuclei: microstructural, mineralization and atrophy modifications measured in vivo using MRI. Neuroimage 2009;48:29-36.
30.
11.
O’Dwyer L, Lamberton F, Bokde AL, et al. Using support vector machines with multiple indices of diffusion for automated classifi-
Cherubini A, Peran P, Spoletini I, et al. Combined volumetry and DTI in subcortical structures of mild cognitive impairment and Alzheimer’s disease patients. J Alzheimers Dis 2010;19:1273-1282.
Movement Disorders, Vol. 29, No. 9, 2014
1219
28.
E T
A L
Schnoebelen S, Semrud-Clikeman M, Pliszka SR. Corpus callosum anatomy in chronically treated and stimulant na€ıve ADHD. J Atten Disord 2010;14:256-266.
Magnetic Resonance Support Vector Machine Discriminates Essential Tremor With Rest Tremor From Tremor-Dominant Parkinson Disease , MD,1† Andrea Cherubini, PhD,1† Rita Nistico Fabiana Novellino, MD,1 Maria Salsone, MD,2 Salvatore Nigro, PhD,1 Giulia Donzuso, MD1 and Aldo Quattrone, MD1,2* 1
Neuroimaging Research Unit, Institute of Molecular Bioimaging and Physiology-National Research Council, Catanzaro, Italy and; 2 Institute of Neurology, Magna Graecia University, Catanzaro, Italy
ABSTRACT Background: The aim of the current study was to distinguish patients who had tremor-dominant Parkinson’s disease (tPD) from those who had essential tremor with rest tremor (rET). Methods: We combined voxel-based morphometryderived gray matter and white matter volumes and diffusion tensor imaging-derived mean diffusivity and fractional anisotropy in a support vector machine (SVM) to evaluate 15 patients with rET and 15 patients with tPD. Dopamine transporter single-photon emission computed tomography imaging was used as ground truth. Results: SVM classification of individual patients showed that no single predictor was able to fully discriminate patients with tPD from those with rET. By contrast, when all predictors were combined in a multimodal algorithm, SVM distinguished patients with rET from those with tPD with an accuracy of 100%. Conclusions: SVM is an operator-independent and automatic technique that may help distinguish patients with C 2014 tPD from those with rET at the individual level. V International Parkinson and Movement Disorder Society Key Words: resting tremor; magnetic resonance imaging; support vector machine; computer-aided diagnosis
therapeutic strategies and have different prognoses. In this context, dopamine transporter single-photon emission computed tomography (DAT-SPECT) plays a major role, yet it represents a highly invasive and costly diagnostic procedure.3 Over the past decade, advanced magnetic resonance imaging (MRI) methodologies have been applied to group-level image analyses with the aim of quantifying the pattern of structural alterations between patients with PD, patients with ET, and healthy controls.4-6 Those studies have shown brain atrophy in many cortical and subcortical regions, particularly in the basal ganglia, in the substantia nigra of patients with PD, and in the cerebellum of patients with ET. However, currently, no data exists on MRI in patients with rET, and only a few studies have focused on MRI in patients with tPD. Some authors observed an increased gray matter (GM) volume in the nucleus ventralis intermedius of the thalamus in patients with tPD compared with healthy controls,7 whereas others reported reduced brain activity in the prefrontal cortex and globus pallidus of patients who had non-tPD compared with both patients who had tPD and healthy controls.8 Recently, support vector machines (SVMs) have been applied with excellent results to the analysis of MRI data for the computer-aided diagnosis of patients with different neurological diseases.9-16 SVMs differ from conventional MRI studies performed at a group level, because they can label each patient using a combination of multiple features extracted from individual images. In the current study, we investigated the potential usefulness of SVMs for differentiating patients with tPD from those with rET using a combined whole-brain, voxel-based morphometry (VBM) and diffusion tensor imaging (DTI) analyses.
Patients and Methods We enrolled 15 patients who had a diagnosis of tPD according to established clinical criteria17 and 15 patients who had rET.18 In patients with tPD and rET, clinical evaluations and tremor scores were calculated according to the Unified Parkinson’s Disease Rating Scale motor examination (UPDRS-ME) (section III, items 20 and 21) and the Fahn-Tolosa scale. No
-----------------------------------------------------------*Correspondence to: Prof. Aldo Quattrone, Institute of Neurology, Magna Graecia University, Viale Europa, 88100, Catanzaro, Italy; [email protected] †
Differentiating between patients who have essential tremor (ET) with resting tremor (rET) from those who have tremor-dominant Parkinson’s disease (tPD) may be challenging in the early phases of the diseases.1,2 Distinguishing between these pathologies, however, is an important task, because they demand different
1216
Movement Disorders, Vol. 29, No. 9, 2014
The first two authors contributed equally to this work.
Relevant conflicts of interest/financial disclosures: Nothing to report. Full financial disclosures and author roles may be found in the online version of this article. Received: 31 July 2013; Revised: 31 January 2014; Accepted: 17 February 2014 Published online 13 April 2014 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/mds.25869
M R - S V M
patient had any history of thyroid disease, cerebrovascular disease, or other neurologic diseases. All participants underwent the same imaging protocol, including MRI and DAT-SPECT. Patients were examined using a 3-Tesla GE MR750 scanner (GE Healthcare, Rahway, NJ). The MRI protocol included whole-brain, three-dimensional, T1-weighted, spoiled gradient recall echo (voxel size 5 1 3 1 3 1 mm3); DTI (b 5 0,1000; 27 directions; voxel size 5 2 3 2 3 2 mm3); and conventional clinical sequences. Image processing was performed using the Functional Magnetic Resonance Imaging of the Brain (FMRIB) Analysis Group Software Library (FSL) (Oxford University, United Kingdom).19 T1-weighted volumes were analyzed using the FSL voxel-based morphometry (FSLVBM) tool, and DTI volumes were analyzed with FSLDTIfit. To compare structural patterns across individuals, each individual image was transformed into a common coordinate system, called standard stereotactic space. The modulated VBM-style segmented images (GM volume and white matter [WM] volume), the mean diffusivity (MD), and fractional anisotropy (FA) maps, all in standard space, represented the four features for each voxel that were fed to the SVM algorithm in order to classify single patients. Images in standard space were smoothed with a Gaussian kernel (6 mm) before statistical processing. Voxel-wise, conventional, univariate statistical analysis on each of the four predictors was performed in RANDOMISE, part of FSL. Age and gender were used as non-explanatory coregressors. The individual-level SVM analysis was performed with LIBSVM20 using the method described by Zhang et al.13 and Mwangi et al.14 for multimodal image markers. Although the SVM is capable of working with a large number of features—in our case, four features per voxel—not all of these will be equally relevant for the classification. Consequently, we preselected the most important brain regions by using feature selection. Machine learning studies have demonstrated that feature selection is important for achieving a high predictive accuracy classifier.21 We performed an F test during training to identify the voxels that differed mostly in rET versus tPD. An optimal threshold of P was identified by empirically varying the threshold using a cross-validation procedure with the training data.14 A multi-kernel approach was employed when all predictors were combined.13 Sensitivity and specificity were assessed by a leaveone-out strategy: each patient, in turn, was removed both from the training set and from the generation of the weighting map. The left-out patient was then classified using the trained model to evaluate the accuracy of the algorithm. Before inclusion in the study, written informed consent was obtained from all participants, and the institutional review board approved the study.
D I S C R I M I N A T E S
B E T W E E N
r E T
A N D
t P D
TABLE 1. Demographic, clinical, and imaging features in patients with tremor-dominant Parkinson’s disease and patients with essential tremor with rest tremor Characteristic
tPD, n 5 15
rET, n 5 15
P
Age: Mean 6 SD, y Men: No. (%) Age at onset: Mean 6 SD, y Duration of disease: Mean 6 SD, y Familial history: No. (%) UPDRS-ME score: Mean 6 SD Fahn-Tolosa score: Mean 6 SD DAT-SPECT:d Mean 6 SD Left putamen Right putamen
68.7 6 5.8 9 (60%) 61.2 6 6.7
68.1 6 8.8 8 (53) 58.3 6 8.7
NSa NSb NSa
7.7 6 3.0
9.9 6 7.1
NSa
4 (22%) 13.0 6 1.9
7 (46) 11.4 6 5.7
NSb NSc
18.9 6 5.4
18.6 6 11.8
NSc
1.04 6 0.16 1.08 6 0.14
2.07 6 0.30 2.15 6 0.30
< 0.001a < 0.001a
a
This P value was value determined using an unpaired t test. This P value was value determined using a v2 test. This P value was value determined using the Mann-Whitney test. Abbreviations: tPD, tremor-dominant Parkinson’s disease; rET, essential tremor with rest tremor; SD, standard deviation; NS, nonsignificant; UPDRS-ME, Unified Parkinson’s Disease Rating Scale-Motor Examination (items 20 and 21); DAT-SPECT, dopamine transporter single-photon emission computed tomography. d The DAT-SPECT value was determined using the ratio of specific to nonspecific radioligand binding. b c
Results Clinical, demographic, and DAT-SPECT characteristics of the tPD and rET groups are shown in Table 1. Patients with rET had normal DAT-SPECT imaging, whereas tracer uptake in patients with tPD was decreased in the putamina. VBM and DTI results from the conventional univariate F test are shown in Figure 1A at a statistical threshold of P < 0.001. No voxel survived correction for multiple comparisons in any of the four parameters considered (GM, WM, MD, and FA). Group differences at the uncorrected threshold between rET and tPD patients were localized in caudate nucleus, globus pallidus, midbrain, internal capsule, body of the corpus callosum, and cerebellum regions. The SVM algorithm showed a high accuracy for distinguishing rET from tPD when GM or WM was used as a single predictor, whereas FA and MD were less accurate. When the four predictors (GM, WM, FA, and MD) were used simultaneously in a multimodal SVM, the algorithm had an accuracy of 100% for distinguishing patients with rET from those with tPD, thus allowing a differential diagnosis at the individual level. A graphic representation of this analysis is provided in Figure 1B.
Discussion In the current study, for the first time, we used a combination of VBM and DTI analyses for the differentiation of patients with rET from those with tPD. Multi-modal pattern recognition of MRI data was
Movement Disorders, Vol. 29, No. 9, 2014
1217
C H E R U B I N I
E T
A L
FIG. 1. The results from a conventional univariate F test between patients who had tremor-dominant Parkinson’s disease (tPD) and patients who had essential tremor with resting tremor (rET) are superimposed on a standard template. (A) Selected axial slices show the significant (P < 0.001, uncorrected) voxels for different predictors, represented by the voxel-based morphometry (VBM)-derived gray matter (GM) and white matter (WM) volume and the diffusion tensor imaging (DTI)-derived mean diffusivity (MD) and fractional anisotropy (FA). (B) Performances of a support vector machine (SVM) classifier were obtained using the four voxel-based predictors. Color is indicative of the diagnosis according to clinical criteria and dopamine transporter single-photon emission computed tomography imaging: red indicates patients with rET, and cyan indicates patients with tPD. Patients are distributed along a vertical axis according to the probability of belonging to one of the two groups resulting from SVM classification. The far right column shows the performance of the classifier using all four predictors combined (All).
capable of fully discriminating between patients with tPD and patients with rET at the individual level. Although a differential diagnosis between rET and tPD is possible using DAT-SPECT imaging, this technique is invasive and expensive as a routine diagnostic tool. To our knowledge, there are currently no studies investigating the potential of SVM for the individual detection of rET, and only few studies have investigated computer-aided diagnosis of PD using MRI.10,12,15,16 Conventional analyses of MRI data have been conducted in patients with ET by studying VBM,22,23 DTI,6,24,25 and iron accumulation26 compared with healthy controls. Although most studies suggested an impairment of the cerebello-thalamocortical loop, the results remain controversial. Regarding rET, currently, there are no MRI studies focused on this clinical subtype of ET. Similarly, studies investigating brain structural alterations in patients with PD have not yet drawn any congruent conclusion. Some studies showed that there was no structural difference between patients with PD and controls, and others showed a nigrostriatal signature of PD and an impairment of the basal gangliathalamocortical loop.4,27,28
1218
Movement Disorders, Vol. 29, No. 9, 2014
Our results of conventional MRI univariate analyses are in agreement with the literature, although the limited sample size of our study did not allow for high statistical significance. Thus, we attempted to increase the reliability of MRI by combining VBM and DTI in an SVM algorithm with the aim of increasing the overall sensitivity of the analysis. Our results showed that the performances of single classifiers (GM, WM, FA, MD) were lower than the multi-modal method classifier (Fig. 1B). Indeed, when the four image predictors were fed simultaneously to the SVM algorithm, we observed an overall accuracy of 100% in distinguishing patients with tPD from those with rET. Several works have highlighted the finding that multimodal MRI is better at capturing the chain of structural alterations acting simultaneously on brain tissue at the macroscopic and microscopic levels.29,30 Indeed, this might be the reason why the best accuracy was observed only when all predictors were used by the SVM at the same time. There are some limitations to this study. First, this was a preliminary study with a small sample size, and the reported accuracy might vary when the method is employed on a larger population with more
M R - S V M
heterogeneous phenotypes. Second, we used the same patients in the training and testing of the SVM algorithm, a procedure that could lead to data over fitting and, consequently, to an overestimation of classification accuracy. To limit this problem, we applied a leaveone-out protocol, ensuring that the classifier was always tested on new data sets, ie, data sets that had not been used previously to build the classification model. In conclusion, SVM analysis of voxel-based MRI is an operator-independent and automatic postprocessing technique that may help to correctly classify patients with resting tremor at the individual level and in the absence of a DAT scan investigation. SVM may be used for the differentiation of patients who have clinically indistinguishable phenotypes, such as those with tPD and rET, thus representing a valid help in everyday clinical practice.
References 1.
Thenganatt MA, Louis ED. Distinguishing essential tremor from Parkinson’s disease: bedside tests and laboratory evaluations. Expert Rev Neurother 2012;12:687-696.
2.
Cohen O, Pullman S, Jurewicz E, Watner D, Louis ED. Rest tremor in patients with essential tremor: prevalence, clinical correlates, and electrophysiologic characteristics. Arch Neurol 2003;60: 405-410.
D I S C R I M I N A T E S
B E T W E E N
r E T
A N D
t P D
cation of mild cognitive impairment [serial online]. PLoS One 2012;7:e32441. 12.
Long D, Wang J, Xuan M, et al. Automatic classification of early Parkinson’s disease with multi-modal MR imaging [serial online]. PLoS One 2012;7:e47714.
13.
Zhang D, Wang Y, Zhou L, Yuan H, Shen D. Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage 2011;55:856-867.
14.
Mwangi B, Ebmeier KP, Matthews K, Steele JD. Multi-centre diagnostic classification of individual structural neuroimaging scans from patients with major depressive disorder. Brain 2012;135: 1508-1521.
15.
Haller S, Badoud S, Nguyen D, Garibotto V, Lovblad KO, Burkhard PR. Individual detection of patients with Parkinson disease using support vector machine analysis of diffusion tensor imaging data: initial results. AJNR Am J Neuroradiol 2012;33: 2123-2128.
16.
Haller S, Badoud S, Nguyen D, Barnaure I, Montandon ML, Lovblad KO, Burkhard PR. Differentiation between Parkinson disease and other forms of Parkinsonism using support vector machine analysis of susceptibility-weighted imaging (SWI): initial results. Eur Radiol 2013;23:12-19.
17.
Gelb DJ, Oliver E, Gilman S. Diagnostic criteria for Parkinson disease. Arch Neurol 1999;56:33-39.
18.
Deuschl G, Bain P, Brin M. Consensus statement of the Movement Disorder Society on tremor.Ad Hoc Scientific Committee. Mov Disord 1998;13(suppl 3):2-23.
19.
Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. Neuroimage 2012;62:782-790.
20.
Chang C-C, Lin C-J. LIBSVM: a library for support vector machines. ACM T Intel Syst Technol 2011;2:3-27.
21.
Guyon I, Elisseeff A. An introduction to variable and feature selection. J Mach Learn Res 2003;3:1157-1182.
3.
Badiavas K, Molyvda E, Iakovou I, Tsolaki M, Psarrakos K, Karatzas N. SPECT imaging evaluation in movement disorders: far beyond visual assessment. Eur J Nucl Med Mol Imaging 2011;38: 764-773.
22.
Benito-Leon J, Alvarez-Linera J, Hernandez-Tamames JA, AlonsoNavarro H, Jimenez-Jimenez FJ, Louis ED. Brain structural changes in essential tremor: voxel-based morphometry at 3-Tesla. J Neurol Sci 2009;287(1–2):138-142.
4.
Stoessl AJ, Martin WW, McKeown MJ, Sossi V. Advances in imaging in Parkinson’s disease. Lancet Neurol 2011;10:987-1001.
23.
5.
Passamonti L, Cerasa A, Quattrone A. Neuroimaging of essential tremor: what is the evidence for cerebellar involvement [serial online]? Tremor Other Hyperkinet Mov (NY) 2012;2:02-67-421-3.
Quattrone A, Cerasa A, Messina D, et al. Essential head tremor is associated with cerebellar vermis atrophy: a volumetric and voxelbased morphometry MR imaging study. AJNR Am J Neuroradiol 2008;29:1692-1697.
24.
6.
Prodoehl J, Li H, Planetta PJ, et al. Diffusion tensor imaging of Parkinson’s disease, atypical parkinsonism, and essential tremor. Mov Disord 2013;28:1816-1822.
Nicoletti G, Manners D, Novellino F, et al. Diffusion tensor MRI changes in cerebellar structures of patients with familial essential tremor. Neurology 2010;74:988-994.
25.
7.
Kassubek J, Juengling FD, Hellwig B, Spreer J, Lucking CH. Thalamic gray matter changes in unilateral Parkinsonian resting tremor: a voxel-based morphometric analysis of 3-dimensional magnetic resonance imaging. Neurosci Lett 2002;323:29-32.
Saini J, Bagepally BS, Bhatt MD, et al. Diffusion tensor imaging: tract based spatial statistics study in essential tremor. Parkinsonism Relat Disord 2012;18:477-482.
26.
Novellino F, Cherubini A, Chiriaco C, et al. Brain iron deposition in essential tremor: a quantitative 3-Tesla magnetic resonance imaging study. Mov Disord 2013;28:196-200.
27.
Peran P, Cherubini A, Assogna F, et al. Magnetic resonance imaging markers of Parkinson’s disease nigrostriatal signature. Brain 2010;133:3423-3433.
8.
Prodoehl J, Planetta PJ, Kurani AS, Comella CL, Corcos DM, Vaillancourt DE. Differences in brain activation between tremorand nontremor-dominant Parkinson disease. JAMA Neurol 2013; 70:100-106.
9.
Kloppel S, Abdulkadir A, Jack CR Jr, Koutsouleris N, MouraoMiranda J, Vemuri P. Diagnostic neuroimaging across diseases. Neuroimage 2012;61:457-463.
28.
Pan PL, Song W, Shang HF. Voxel-wise meta-analysis of gray matter abnormalities in idiopathic Parkinson’s disease. Eur J Neurol 2012;19:199-206.
10.
Cherubini A, Morelli M, Nistico R, et al. Magnetic resonance support vector machine discriminates between Parkinson disease and progressive supranuclear palsy [published online ahead of print 3 Dec 2013]. Mov Disord 2013. doi: 10.1002/mds.25737.
29.
Cherubini A, Peran P, Caltagirone C, et al. Aging of subcortical nuclei: microstructural, mineralization and atrophy modifications measured in vivo using MRI. Neuroimage 2009;48:29-36.
30.
11.
O’Dwyer L, Lamberton F, Bokde AL, et al. Using support vector machines with multiple indices of diffusion for automated classifi-
Cherubini A, Peran P, Spoletini I, et al. Combined volumetry and DTI in subcortical structures of mild cognitive impairment and Alzheimer’s disease patients. J Alzheimers Dis 2010;19:1273-1282.
Movement Disorders, Vol. 29, No. 9, 2014
1219
