Brain morphometry of MR images for automated classification of first-episode schizophrenia

Information Fusion xxx (2013) xxx–xxx

Contents lists available at SciVerse ScienceDirect

Information Fusion journal homepage: www.elsevier.com/locate/inffus

Brain morphometry of MR images for automated classification of first-episode schizophrenia Daniel Schwarz a,⇑, Tomas Kasparek b,c a

Institute of Biostatistics and Analyses, Masaryk University, Kamenice 5, 625 00 Brno, Czech Republic Behavioural and Social Neuroscience Group, CEITEC – Central European Institute of Technology, Masaryk University, Czech Republic c Department of Psychiatry, University Hospital Brno and Masaryk University, Jihlavska 20, 625 00 Brno, Czech Republic b

a r t i c l e

i n f o

Article history: Received 8 March 2012 Received in revised form 9 September 2012 Accepted 4 February 2013 Available online xxxx Keywords: Image registration Brain morphometry Pattern recognition Computational neuroanatomy Group analysis Schizophrenia

a b s t r a c t Schizophrenia is a disabling psychiatric disorder that usually begins to affect individuals during their adolescence or early adulthood and most patients continue to suffer social, economic, and psychological difficulties from the very onset of the disorder. The neurobiology of the disorder comprises of changes in the brain that can be detected using MR imaging. Focus on the morphological changes in patients with first episode schizophrenia limits the confounding effect of factors such as long-term medication or progression. Therefore, the detected abnormalities are more likely to indicate the primary pathology and the existence of such changes provides the opportunity of applying them for subject recognition based on brain imaging. This paper presents a combination of methods pertaining to automated whole-brain morphometry of MR images and the methods of supervised learning. The designed recognition procedure is successfully used here for classification of 104 subjects into groups of patients and healthy volunteers with the use of k-NN and SVM classifiers. The same algorithm is further used for distinguishing between patients who responded well to treatment and those who did not show adequate symptomatic relief. Ó 2013 Elsevier B.V. All rights reserved.

1. Introduction The rising trend of diagnostic medical imaging has given great significance to new methods of image data analyses. Automated whole-brain morphometry, in its variants, has been extensively applied in the last decades alongside volumetric based analysis of regions of interest (ROI) on data from structural brain magnetic resonance imaging (MRI). While the common ROI methods requires manual segmentation, which is very time consuming and prone to subjective errors, whole-brain methods apply fully automatic processing of images, including their registration and segmentation. An advantage of whole-brain automated morphometric methods is that the detection of pathological changes in the brain morphology is not limited by a priori determined boundaries or regions, which is very beneficial in neuropsychiatric disorders such as schizophrenia [1,2]. In addition to the morphometric analyses, MR image data has been recently also used for uncovering spatially complex patterns which distinguish patients suffering from neuropsychiatric disorders from healthy volunteers. Assuming that the classification algorithms were of high accuracy, MRI-based prediction of a neu⇑ Corresponding author. E-mail addresses: [email protected] (D. Schwarz), [email protected] (T. Kasparek).

ropsychiatric disorder would be possible at individual level. Such classification accuracy has already achieved clinically meaningful values in Alzheimer’s dementia – over 95% correctly classified subjects simply on the basis of features deduced from structural brain images [3]. Alzheimer’s dementia is, however, accompanied with significant changes in the anatomy. In case of disorders with less prominent anatomical changes – such as schizophrenia – the classification accuracy is between 70% and 90% in most of the studies [4–8]. Of special interest are the cases of early manifestation of the disorder where timely detection of the subjects at risk of developing schizophrenia [9] or subjects with first-episode [5] represents an urgent clinical priority. Till today, several pre-processing approaches have been applied for classification purposes: MR image segmentation was involved [10,11] where the feature vectors were formed by tissue densities for GM, WM and CSF, and a nonlinear support vector machine (SVM) classifier was then constructed from the data of schizophrenia patients and healthy controls. Another classification procedure [12] derived the features from parameters of optimal affine transform which mapped selected anatomical structures between the images and a digital brain atlas. The authors reported a leave-one-out cross-validation accuracy of up to 90% but without the sensitivity and specificity values; moreover, this result was achieved on a very small number of training patterns. Similar problematic results are analyzed in another review [13], which warns against common errors in reporting

1566-2535/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.inffus.2013.02.002

Please cite this article in press as: D. Schwarz, T. Kasparek, Brain morphometry of MR images for automated classification of first-episode schizophrenia, Informat. Fusion (2013), http://dx.doi.org/10.1016/j.inffus.2013.02.002

2

D. Schwarz, T. Kasparek / Information Fusion xxx (2013) xxx–xxx

prediction accuracies for various numbers of training patterns in each class and the poorly chosen validation techniques. All of these studies used one modality of information on brain pathology for classification purposes. There is evidence of relationship between the magnitude of information provided and the classification accuracy [13,14]. It would, therefore, be advantageous to use a set of various brain morphometric parameters to increase the accuracy of the classification algorithms. In case of automated morphometric analysis, there are two methods that provide different information about brain morphology: voxel-based morphometry (VBM) and deformation-based morphometry (DBM) providing information on the local gray matter volume and the local brain volume changes respectively. However, the magnitude of information that can be put to practical use is limited by the multidimensionality of the brain images – therefore, a certain extent of dimensionality reduction is a necessary prerequisite for classification analyses. This paper presents a procedure designed for recognizing patients with first-episode schizophrenia from healthy volunteers based on their structural T1-weighted MR images. The features are extracted from the results of two different automated wholebrain morphometric methods in a way that keeps in balance the need for data dimensionality reduction and the richness of information provided on brain morphology. SVM and k-NN classifiers are constructed from the features of 52 first-episode patients and 52 normal controls. The same recognition procedure is then applied on the problem of distinguishing between schizophrenia patients who responded well to the instituted acute treatment, from the patients who had no or poor improvement of the symptoms. 2. Methods The proposed MRI-based procedure for recognizing schizophrenia patients and for predicting the outcome of the acute treatment is shown in Fig. 1. The image pre-processing phase includes the VBM and DBM methods, each employing a different registration algorithm. The SVM and k-NN classifiers are trained and tested on the features derived from each morphometric method separately as well as on the features extracted from significant results of both morphometric methods. 2.1. Study design We performed a prospective observational study of patients with schizophrenia during their first episode to assess the usefulness of the MRI morphometric examination for diagnostic and treatment outcome prediction purposes. 2.2. Subjects 2.2.1. Schizophrenia patients Fifty-two patients (mean age 24, SD 5.1 years) admitted to the all-male unit of the Department of Psychiatry, Masaryk University in Brno, for first episode of schizophrenia (that is their symptoms fulfilled the criteria for schizophrenia for the first time when admitted to the department, including the time criterion – duration of symptoms longer than 1 month) were recruited. Diagnosis was established during clinical interviews held in compliance with the International Statistical Classification of Disease and Related Health Problems (ICD-10) research criteria. The diagnosis was established as follows: the subject underwent a clinical interview focused on information about the family and personal history, somatic conditions, substance abuse, pharmacological history and current treatment, previous psychiatric conditions and, finally, on the current clinical manifestation, the presenting

Fig. 1. Training phase of the recognition algorithm. Feature vectors for classification of schizophrenia patients and normal controls are obtained by selecting the most statistically significant local volume changes and gray matter densities detected with VBM and DBM. The same features are further used for predicting reduction in schizophrenia symptoms.

symptoms, duration, and the functional impact. The interview was performed by a psychiatrist trained in conducting diagnostic interviews. Where possible, information from relatives was collected as well. Next, the patients were physically examined, inclusive of laboratory tests for blood (hematology, biochemistry) and urine analysis (biochemistry, toxicology). Abnormal findings were subjected to additional examinations and tests. A fully trained senior psychiatrist reviewed all information, established the diagnosis and suggested the case for inclusion in the study. An investigating senior psychiatrist again reviewed the clinical information and the diagnosis, and checked for exclusion criteria. 87% of the patients suffered from Paranoid schizophrenia. The course of the treatment and prescribing the dose followed the general clinical recommendations — psychotropic medication included antipsychotics, benzodiazepines, hypnotics, and anticholinergic antiparkinsonics. All patients were treated with atypical antipsychotics. Symptomatology was evaluated using the Positive and Negative Syndrome Scale (PANSS [15]) at admission and at the end of acute treatment. The end of the acute phase was defined as having the Clinical Global Impression-Severity score of below 3. Patients were rated every week following admission. The magnitude of the total PANSS score reduction during treatment of the first episode was calculated and applied in prediction analysis. Exclusion criteria pertained to substance dependence detected by clinical evaluation and urine toxicology tests, neurological or systemic disease with known relationship to brain alteration detected by clinical evaluation, physical and neurological examinations, serum and urine chemistry and blood count, serological examination for neurotropic agents, clinical evaluation of MRI scans, and contraindications for MRI.

Please cite this article in press as: D. Schwarz, T. Kasparek, Brain morphometry of MR images for automated classification of first-episode schizophrenia, Informat. Fusion (2013), http://dx.doi.org/10.1016/j.inffus.2013.02.002

D. Schwarz, T. Kasparek / Information Fusion xxx (2013) xxx–xxx

2.2.2. Healthy controls Fifty-two healthy subjects (matched for age – mean age 24, SD 3.7 years – gender and handedness) were recruited from the community, the local staff, and medical students. The exclusion criteria (as assessed during clinical interviews performed by a trained psychiatrist) were substance dependence, family history of axis I psychiatric conditions, personal history of axis I psychiatric condition, neurological or somatic conditions affecting the structure or function of the brain, and contraindications for MRI examination. The study was approved by the local ethics committee and all subjects signed the informed consent. 2.3. Image pre-processing and spatial normalization The dataset contained 104 T1-weighted images of the entire head obtained with the 1.5 T MR device. Sagittal tomographic plane thickness was 1.17 mm, the in-plane resolution was 0.48 mm  0.48 mm (3-D field of view contained 160  512  512 voxels) and image intensities were represented by 12-bit integers. MRI examination was performed at the end of the first episode. GM tissue segments were obtained from all images after correction for bias-field inhomogeneity, spatial normalization and segmentation [16] with the use of VBM8 toolbox (http:// dbm.neuro.uni-jena.de/vbm/) implemented in SPM8 framework (http://www.fil.ion.ucl.ac.uk/spm/). Spatial normalization steps involved affine registration to standard SPM T1 template followed by fast diffeomorphic registration algorithm DARTEL [17]. GM tissue segments were modulated with the determinant of Jacobian matrices of the deformations to account for registration related changes in local volumes. The modulated GM segment images were finally smoothed with Gaussian low-pass filter to enable intersubject comparisons and to render the data distribution more normal. A wide range of smoothing kernels has been used in VBM studies. Jones et al. investigated in [18] the effect of filter size on the results of VBM analyses and stated that the kernel size should be matched to the spatial extent of the expected differences between studied groups. The 8 mm FWHM Gaussian kernel, which was used here, complied with that rule. Spatial normalization steps for DBM included the same affine registration algorithm as for VBM. After transforming all bias-corrected images into stereotactic space, our original high-dimensional deformable registration technique was used to compute vector displacement fields which maximized the normalized mutual information between the images and the high-resolution single-subject ICBM (International Consortium for Brain Mapping) template. The registration algorithm involved calculation of local forces in each voxel and their regularization with the use of modified Rogelj’s elastic-incremental spatial deformation model; for more details see [19,20]. The resulting 3-D displacement vector fields were converted into scalar fields by computing their respective determinants of Jacobian matrices at each voxel of the stereotactic space. 2.4. Feature selection and classification The modulated and smoothed images from VBM reflect spatial distribution of GM density. Voxel-wise two-sample t-tests of GM density means in FES and NC groups with the subject’s age as single covariate were used to compute the distribution of Student’s tstatistics in the stereotactic space. The map of t-statistic was thresholded at the significance level of p < 0.001 without correction for multiple comparisons, consequent to which voxel clusters less than 100 mm3 were filtered out. The resulting binary mask was used to select the classification features. These features, extracted with the use of VBM, in fact represent statistically significant

3

reduction in GM density which might be interpreted, for a brain region localized by a cluster of significant voxels, as loss of GM tissue volume in FES patients when compared to NC subjects. Scalar fields from DBM represented volume reduction of the voxels in which the Jacobian determinant was less than one, and volume expansion of the voxels in which the Jacobian determinant was greater than one. The data was log-transformed and normally distributed reductions (negative values) and expansions (positive values) were obtained. Statistical analysis, as in the case of VBM, was used to detect significant differences between the FES and NC groups. The resulting binary mask was used to select the second set of features for classification. These DBM features reflect significant local volume changes in FES patients when compared to NC subjects. Two well-known methods, namely k-nearest neighbors (k-NN) and support vector machines (SVMs) with linear kernel are compared on the above described dataset. SVM classifier was chosen because of its good performance reported in MRI data classification [8–10,12]. The k-NN algorithm was chosen as a very simple and one of the most popular method in machine learning, which may outperform SVM in selected tasks [21]. The same three training sets of feature vectors were used in both classification algorithms: (1) GM densities obtained from VBM, (2) local volume changes obtained from DBM, and (3) union of the two feature sets. Since two morphometric parameters were available, classification with the use of each separately was not performed, as this could lead to indeterminate results. The quality of classification was determined in a leave-one-out cross validation procedure on 104 subjects from FES and NC groups. For each experiment 103 subjects were used for training and the remaining one for testing. Sensitivity, specificity, precision, overall accuracy and error were assessed based on classification of testing subjects:

Sensitivity ¼

TP ; TP þ FN

ð1Þ

Specificity ¼

TN ; TN þ FP

ð2Þ

Accuracy ¼

TP þ TN ; TP þ TN þ FP þ FN

ð3Þ

Precision ¼

TP ; TP þ FP

ð4Þ

Error ¼

FP þ FN ; TP þ TN þ FP þ FN

ð5Þ

where TP refers to true positive results, TN denotes true negative results, FP denotes false positive results and FN denotes false negative results. The same sets of features, but only for 52 FES patients, were further used for predicting the outcome of the acute treatment instituted for the first schizophrenia episode. Positive and Negative Syndrome Scale (PANSS) was used for rating the symptoms of schizophrenia. A reduction in the initial PANSS total score was used here as cut-off to define two subgroups of FES patients. The group with higher percentage PANSS total score reduction contained FES patients who responded well to the treatment and the group with lower percentage PANSS score reduction contained FES patients with no or poor improvement. The interval of all the PANSS score reduction values was sampled 50 times equidistantly and the true positive rate as well as the false positive rate was evaluated for each cut-off.

Please cite this article in press as: D. Schwarz, T. Kasparek, Brain morphometry of MR images for automated classification of first-episode schizophrenia, Informat. Fusion (2013), http://dx.doi.org/10.1016/j.inffus.2013.02.002

4

D. Schwarz, T. Kasparek / Information Fusion xxx (2013) xxx–xxx

– Euclidean and Cosine – and for different number of neighbors k. Complete NC-FES classification results including sensitivity, specificity, precision and error in selected classifiers only are shown in Table 1. In the prediction experiment, the interval of the PANSS score reduction in 52 FES subjects acquired values of TPchange 2 h0.990; +0.097i. The true positive rates and the false positive rates of the SVM and 11-NN classifiers were computed for all TPchange cut-offs and for all three cases – with the use of both sets of features individually as well as with their union – see Fig. 3. 4. Discussion

Fig. 2. Accuracy of FES-NC classification using gray matter density features (GM) and with the use of features based on local volume changes (VOL).

3. Results VBM selected 14,700 significant GM densities and DBM selected local volume changes in 11,568 significant voxels. The united feature set was represented by the total number of 26,268 features. The description of each subject was reduced to 1.36% of the data when compared to all 1,924,670 voxels covered by the brain tissues of one subject in the stereotactic space. The quality of k-NN and SVM classification of FES and NC subjects using both sets of features as well as their union is expressed in Fig. 2 in terms of accuracy. In the case of k-NN classifier, the cross validation procedure was run with various distance metrics

Fig. 2 and Table 1 show that recognition based on the united set of features selected from voxel-based and deformation-based morphometry provides better results – sensitivity 88.5% and specificity 86.5% – than in the case of using only features selected with one morphometric method. Therefore, the more information on brain morphology we use for classification purposes, the better the classifier performance. One might assume that these well discriminating features for FES-NC recognition selected on the basis of group-level analyses will be closely linked to the pathology of schizophrenia, and thus might well discriminate the groups of patients with different outcomes. This assumption is confirmed here only in the case of features selected by VBM, with which the linear-kernel SVM classifier predicted 74.6% reduction in PANSS total score with sensitivity and specificity of 87.1% and 66.7% respectively. Given that fact such large reduction occurred in 21 out of 52 FES patients in the investigated group, we consider the result as worthy of note; although clinical studies focused on evaluating the effect of treatment operate with PANSS total score reduction cut-off of 30% and 50% [22]. Surprisingly, the recognition procedure based only on the local volume changes, which seemed to discriminate FES and NC groups almost as well as the union of the two feature sets, gave the worst results in the prediction of treatment outcome, see Figs. 2 and 3.

Fig. 3. The false positive rates (FPRs) and the true positive rates (TPRs) of predictions with (a) SVM classifier and (b) k-NN classifier with 11 neighbors. The rates were obtained for three different sets of features and for various cut-offs in reduction in PANSS total score. The TPchange cut-offs were used to divide 52 FES subjects into two groups. The fitted curves are shown only for better orientation in the x–y plots.

Please cite this article in press as: D. Schwarz, T. Kasparek, Brain morphometry of MR images for automated classification of first-episode schizophrenia, Informat. Fusion (2013), http://dx.doi.org/10.1016/j.inffus.2013.02.002

5

D. Schwarz, T. Kasparek / Information Fusion xxx (2013) xxx–xxx

Table 1 Comparison of selected classification results obtained with the use of gray matter density features (GM) and with features based on local volume changes (VOL). Leave-one-out cross validation procedure was executed on 104 subjects from FES and NC classes.

GM SVM linear 11-NN euclidean 21-NN euclidean VOL SVM linear 11-NN euclidean 21-NN euclidean S GM VOL SVM linear 11-NN euclidean 21-NN euclidean

Accuracy (%)

Sensitivity (%)

Specificity (%)

Precision (%)

Error (%)

65.4 74.0 77.9

65.4 76.9 84.6

65.4 71.2 71.2

65.4 72.7 74.6

34.6 26.0 22.1

76.0 82.7 84.6

73.1 78.8 76.9

78.8 86.5 92.3

77.6 85.4 90.9

24.0 17.3 15.4

83.7 87.5 86.5

84.6 88.5 84.6

82.7 86.5 88.5

83.0 86.8 88.0

16.3 12.5 13.5

Our method is similarly to [10,11] based on a combination of automated whole-brain morphometric methods and pattern recognition algorithms. VBM for all three brain tissues was employed in [10,11] for the feature extraction and local volume changes were then characterized by three variables reflecting density of the tissues. Our approach to local volume changes is based on Jacobian determinants computed directly from the spatial transformations resulting from the high-dimensional deformable registration. The analyzed parameter (change of local volume) has a clear biological meaning. On the other hand, in VBM the meaning of tissue density is much less evident. Our selection of features is performed only with statistical significance thresholding. We believe that this very naive approach might be less prone to overfitting than selection of features with the highest discrimination power derived from correlation with the classification outcome as in [10,11]. We tested the performance of our algorithm on the dataset of images on first-episode schizophrenia patients and healthy controls, whereas chronic schizophrenia patients were also included in the studies [10,11], which showed higher sensitivity and specificity of classification than in our case. Chronic schizophrenia might be, due to the combined effects of the pharmacotherapy and the long-term disease, accompanied with greater changes in the brain morphology than in the early stages of the disease. Only male patients were included to the study; this may decrease the power to generalize the results to all schizophrenia patients. On the other hand, the sample studied was more homogeneous which may increase the sensitivity of the approach. Different templates were used during the spatial normalization in VBM and DBM. A single-subject template with a resolution of 1.0  1.0  1.0 mm was used in our deformable registration technique, whereas the template provided with the VBM8 toolbox was based on 550 healthy adult controls and its resolution was of 1.5  1.5  1.5 mm. Although it is rather a matter of implementation, the use of different templates may introduce an additional degree of freedom. In our case, we preferred, however, to obtain more precise results from DBM rather than to use the same template with the coarser resolution. Usefulness of the whole-brain automated morphometry methods in research on the neurobiology of schizophrenia and other neuropsychiatric diseases has been demonstrated repeatedly by many authors [1,19,23–26]. Here, however, we did not want to deal with uncovering specific regions with significant differences in brain morphology and neither was our goal to compare the findings made using VBM and DBM, as it was explored in [27]. We investigated the possibility to combine the results from the morphometry methods, in order to achieve accurate classification of subjects into diagnostic categories using the image data only. Once a sufficient accuracy is achieved, the imaging methods will be

brought closer to the clinical practice in psychiatry. Further, it will allow an objective diagnosis in psychiatry, which previously was not possible, because current practice is based on subjective assessment of clinically evident changes in mental function and behavior. 5. Conclusions The proposed method, combining the two methods of automated whole-brain morphometry with nonparametric classifiers, successfully recognized patients with first episode schizophrenia from the healthy volunteers. The achieved sensitivity and specificity of 88.5% and 86.5% respectively is comparable to other state-ofthe-art works on MRI-based schizophrenia classification, even though the features were selected in a straightforward process based on simple thresholding of univariate statistics. Furthermore, the method was tested for predicting reduction in schizophrenia symptoms pursuant to acute treatment of the first episode. The leave-one-out cross validation of the prediction method gave reasonable results for reduction in PANSS total score of 74.6%, with sensitivity and specificity at 87.1% and 66.7% respectively. Acknowledgments This work was supported by the Project ‘‘CEITEC – Central European Institute of Technology’’ (CZ.1.05/1.1.00/02.0068) from the European Regional Development Fund and by the research Grant from the Ministry of Health, Czech Republic No. 13359-4. References [1] T. Kasparek, R. Prikryl, M. Mikl, D. Schwarz, E. Ceskova, P. Krupa, Prefrontal but not temporal grey matter changes in males with first-episode schizophrenia, Progress in Neuro-Psychopharmacology and Biological Psychiatry 31 (2007) 151–157. [2] C.K. Chan, X. Di, G.M. McAlonan, Q. Gong, Brain anatomical abnormalities in high-risk individuals, first-episode, and chronic schizophrenia: an activation likelihood estimation meta-analysis of illness progression, Schizophrenia Bulletin 37 (2011) 177–188. [3] C.E. Thomaz, F.L. Duran, G.F. Busatto, D.F. Gillies, D. Rueckert, Multivariate statistical differences of MRI samples of the human brain, Journal of Mathematical Imaging and Vision 29 (2007) 95–106. [4] Y. Kawasaki, M. Suzuki, F. Kherif, T. Takahashi, S.-Y. Zhou, K. Nakamura, et al., Multivariate voxel-based morphometry successfully differentiates schizophrenia patients from healthy controls, NeuroImage 34 (2007) 235–242. [5] T. Kasparek, C.E. Thomaz, J.R. Sato, D. Schwarz, E. Janousova, R. Marecek, et al., Maximum-uncertainty linear discrimination analysis of first-episode schizophrenia subjects, Psychiatry Research: Neuroimaging 191 (2011) 174– 181. [6] C.M. Leonard, J.M. Kuldau, J.I. Breier, P.A. Zuffante, E.R. Gautier, D.-C. Heron, et al., Cumulative effect of anatomical risk factors for schizophrenia: an MRI study, Biological Psychiatry 46 (1999) 374–382.

Please cite this article in press as: D. Schwarz, T. Kasparek, Brain morphometry of MR images for automated classification of first-episode schizophrenia, Informat. Fusion (2013), http://dx.doi.org/10.1016/j.inffus.2013.02.002

6

D. Schwarz, T. Kasparek / Information Fusion xxx (2013) xxx–xxx

[7] K. Nakamura, Y. Kawasaki, M. Suzuki, H. Hagino, K. Kurokawa, T. Takahashi, et al., Multiple structural brain measures obtained by three-dimensional magnetic resonance imaging to distinguish between schizophrenia patients and normal subjects, Schizophrenia Bulletin 30 (2004) 393–404. [8] C. Davatzikos, D. Shen, R.C. Gur, X. Wu, D. Liu, Y. Fan, et al., Whole-brain morphometric study of schizophrenia revealing a spatially complex set of focal abnormalities, Archives of General Psychiatry 62 (2005) 1218–1227. [9] N. Koutsouleris, E.M. Meisenzahl, C. Davatzikos, R. Bottlender, T. Frodl, J. Scheuerecker, et al., Use of neuroanatomical pattern classification to identify subjects in at-risk mental states of psychosis and predict disease transition, Archives of General Psychiatry 66 (2009) 700–712. [10] Y. Fan, D. Shen, R.C. Gur, R.E. Gur, C. Davatzikos, COMPARE: classification of morphological patterns using adaptive regional elements, IEEE Transactions On Medical Imaging 26 (2007) 93–105. [11] Y. Fan, R.E. Gur, R.C. Gur, X. Wu, D. Shen, M.E. Calkins, et al., Unaffected family members and schizophrenia patients share brain structure patterns: a highdimensional pattern classification study, Biological Psychiatry 63 (2008) 118– 124. [12] K. Pohl, M. Sabuncu, A unified framework for MR based disease classification, in: J. Prince, D. Pham, K. Myers (Eds.), Information Processing in Medical Imaging, Springer, Berlin/Heidelberg, 2009, pp. 300–313. [13] O. Demirci, V. Clark, V. Magnotta, N. Andreasen, J. Lauriello, K. Kiehl, et al., A review of challenges in the use of fMRI for disease classification/ characterization and a projection pursuit application from a multi-site fMRI schizophrenia study, Brain Imaging and Behavior 2 (2008) 207–226. [14] F. Shi, Y. Liu, T. Jiang, Y. Zhou, W. Zhu, J. Jiang, et al., Regional homogeneity and anatomical parcellation for fMRI image classification: application to schizophrenia and normal controls, in: Proceedings of the 10th International Conference on Medical Image Computing and Computer-assisted Intervention, Springer-Verlag, Berlin, Heidelberg, 2007, pp. 136–143. [15] S.R. Kay, A. Fiszbein, L.A. Opler, The positive and negative syndrome scale (PANSS) for schizophrenia, Schizophrenia Bulletin 13 (1987) 261–276.

[16] J. Ashburner, K.J. Friston, Unified segmentation, NeuroImage 26 (2005) 839–851. [17] J. Ashburner, A fast diffeomorphic image registration algorithm, NeuroImage 38 (2007) 95–113. [18] D.K. Jones, M.R. Symms, M. Cercignani, R.J. Howard, The effect of filter size on VBM analyses of DT-MRI data, NeuroImage 26 (2005) 546–554. [19] D. Schwarz, T. Kasparek, I. Provaznik, J. Jarkovsky, A deformable registration method for automated morphometry of MRI brain images in neuropsychiatric research, IEEE Transactions On Medical Imaging 26 (2007) 452–461. [20] P. Rogelj, S. Kovacˇicˇ, J.C. Gee, Point similarity measures for non-rigid registration of multi-modal data, Computer Vision and Image Understanding 92 (2003) 112–140. [21] F. Colas, P. Brazdil, Comparison of SVM and some older classification algorithms in text classification tasks, in: M. Bramer (Ed.), Artificial Intelligence in Theory and Practice, Springer US, n.d., pp. 169–178. [22] S. Leucht, J.M. Davis, R.R. Engel, J.M. Kane, S. Wagenpfeil, Defining response in antipsychotic drug trials: recommendations for the use of scale-derived cutoffs, Neuropsychopharmacology 32 (2007) 1903–1910. [23] A. Mechelli, C.J. Price, K.J. Friston, J. Ashburner, Voxel-based morphometry of the human brain: methods and applications, Current Medical Imaging Reviews 1 (2005) 105–113 (9). [24] J. Ashburner, C. Hutton, R. Frackowiak, I. Johnsrude, C. Price, K. Friston, Identifying global anatomical differences: deformation-based morphometry, Human Brain Mapping 6 (1998) 348–357. [25] C.D. Good, I.S. Johnsrude, J. Ashburner, R.N.A. Henson, K.J. Friston, R.S.J. Frackowiak, A voxel-based morphometric study of ageing in 465 normal adult human brains, NeuroImage 14 (2001) 21–36. [26] C. Gaser, I. Nenadic, B.R. Buchsbaum, E.A. Hazlett, M.S. Buchsbaum, Ventricular enlargement in schizophrenia related to volume reduction of the thalamus, striatum, and superior temporal cortex, American Journal of Psychiatry 161 (2004) 154–156. [27] D. Schwarz, T. Kasparek, Comparison of two methods for automatic brain morphometry analysis, Radioengineering 20 (2011) 996–1001.

Please cite this article in press as: D. Schwarz, T. Kasparek, Brain morphometry of MR images for automated classification of first-episode schizophrenia, Informat. Fusion (2013), http://dx.doi.org/10.1016/j.inffus.2013.02.002