Brain Morphometry Methods for Feature Extraction in Random Subspace Ensemble Neural Network Classification of First-Episode Schizophrenia

Schizophrenia (SZ) is a severe mental disorder that affects people most often in early adulthood (Andreasen, 1995). It is characterized by hallucinations and delusions, which impose a profound psychological burden on the patients, their surroundings, and caregivers. If the patient is left untreated for a longer period of time, psychosis is associated with lower levels of symptomatic and functional recovery from the first psychotic episode and also corresponds to the severity of negative symptoms (Perkins, Gu, Boteva, & Lieberman, 2005). Thus, early detection of schizophrenia and administration of antipsychotics is critical.

Computer science is a progressive discipline enabling auxiliary diagnostics through a wide range of tools, such as high-resolution neuroimaging devices and high-speed computers, that are able to run computationally demanding algorithms for image processing, analysis, and pattern recognition. Many papers have revealed the morphological changes in the brain affected by schizophrenia versus the healthy brain, as in Gaser, Volz, Kiebel, Riehemann, and Sauer (1999) and Wright et al. (1995). These differences may enable development of a model applicable in computer-aided diagnostics.

Analysis of brain morphology is an important step on the way to creating a classification framework. The underlying computational neuroanatomy methods, such as voxel- and deformation-based morphometry (Ashburner & Friston, 2000; Ashburner et al., 1998) are widely used for spatial normalization of brain images and detection of morphological abnormalities based on comparisons made between patients and healthy controls.

Voxel-based morphometry (VBM) involves several image processing steps. The first step is spatial normalization of images, that is, registration to a standard template followed by resampling to isotropic voxels and resolution typically set to 1.5 mm or 1.0 mm. This step ensures that global differences in position, orientation, size, and shape are removed while maintaining local differences. The registered images are segmented into tissue types—white matter, gray matter, and cerebrospinal fluid—that are subsequently smoothed by the gaussian filter, and the resulting tissue densities are analyzed statistically. The significant differences identified between tissue densities in healthy and diseased subjects are interpreted as the impact of the disease (Schwarz & Kašpárek, 2011). Following the spatial normalization and segmentation, a step called “modulation” may be incorporated in order to scale the normalized tissue maps by the macroscopic deformations, and thus preserve local volume.

The VBM approach has been validated several times, showing consistent findings obtained from VBM and by means of volumetric calculations over regions of interest (Giuliani, Calhoun, Pearlson, Francis, & Buchanan, 2005; Gong et al., 2005; Keller et al., 2002). However, the underlying methodology of VBM was criticized for being susceptible to errors and false-positive results due to imprecise and possibly erroneous image registrations (Bookstein, 2001) and, subsequently, argued and advocated for (Ashburner & Friston, 2001; Davatzikos, 2004). Criticism against VBM is also alleviated if the SPM package (Statistical Parametric Mapping toolbox for Matlab: http://www.fil.ion.ucl.ac.uk/spm/) is used. The package offers the standard VBM implementation, which includes a precise intersubject registration algorithm DARTEL (diffeomorphic anatomical registration through exponentiated lie algebra) (Ashburner, 2007). Nevertheless, the heterogeneity of the results obtained from VBM analyses is still very high, and the neuroimaging community lacks a gold-standard configuration applicable to all image processing steps within the VBM pipeline, including the choice of a registration algorithm, inclusion of the modulation step, and setting of the gaussian filter smoothing.

The other method for detection of morphological abnormalities presented in this letter is deformation-based morphometry (DBM). The method is based on the analysis of deformation fields obtained from the registration step (Schwarz & Kašpárek, 2011). Unlike VBM, DBM can detect differences in shape and volume within the whole brain. It must be noted that the DBM was originally used (Ashburner et al., 1998) as a method for detecting global brain shape differences in different populations. Several DBM methods differ in the registration algorithm and spatial deformation model. In the early studies (Ashburner & Friston, 2000; Ashburner et al., 1998), smooth parametric transformations with low-frequency sine basis functions were used. This approach (low-resolution DBM) did not take account of all anatomical variability, and therefore it was unable to encode all subtle differences into spatial transformations. Introduction of high-resolution deformable registration has made it possible to describe the complex brain morphology. This high-resolution DBM includes spatial deformation models based on high-dimension parametric transformations or models inspired by similarity to continuum mechanics. There are several ways to analyze the resulting deformation fields. One of the frequently used approaches to the analysis is based on independent univariate tests applied voxel-wise to Jacobian determinants, which quantify changes in voxel volume and can be calculated directly from the deformation fields. In this manner, significant local volume changes between diseased and healthy brains can be detected (Schwarz & Kašpárek, 2011).

Recognizing a mental disease from imaging data, particularly first-episode schiozophrenia, is a complex task employing multivariate algorithms that unveil patterns of subtle differences in the images and use them to assign a class label. Machine learning (ML), which represents the state-of-the-art methodology in the field, encompasses self-adaptive classification algorithms, including, for example, support vector machines (SVM), artificial neural networks (ANN), decision trees, and clustering methods. Promising results from several studies have already been published, showing that ML classifier accuracy can exceed 70%, that is, it is higher than a random guess (accuracy rates are in the brackets): ANNs based on diffusion imaging data (100%; Charpentier, & Savio, 2010), functional magnetic resonance imaging (fMRI) data and features from independent component analysis (75.6%; Jafri & Calhoun, 2006); SVM with wavelet features extracted from MRI (73.20%; Dluhoš, Schwarz, & Kašpárek, 2014); and search-light-based feature extraction from fMRI (91%; Bleich-Cohen et al., 2014). Nieuwenhuis et al. (2012) used SVM and features selected from a gray matter (GM) densities and removed those with low weights in the SVM model (less than 70%). Neural networks also reveal good accuracy for other brain diseases, including Alzheimer's disease (AD): backpropagation network based on MRI and principal component analysis (100%; Huang, Yan, Jiang, & Wang, 2008), radial basis function network, probabilistic neural networks, and learning vector quantization network based on MRI (66–83%; Savio, García-Sebastián, Hernández, Graña, & Villanúa, 2009). Recognition tools for Parkinson's disease include SVM in combination with principal component analysis and MRI (more than 90%, depending on the number of components used; Salvatore et al., 2014).

Ensemble learning represents a methodology that generally employs multiple classifiers, which either differ in principle or share the same principle but differ in some parameter values or are identical but trained on different subsets of data. The underlying algorithms combine individual classifiers' outputs to reach a final decision and thus mimic the thought process of an ensemble. The ensemble learning methods have already been applied in the detection of schizophrenia based on brain imaging data. Yang, Liu, Sui, Pearlson, and Calhoun (2010) achieved classification accuracy of 87% with SVMs governed by the AdaBoost algorithm. Janousova, Schwarz, and Kasparek (2015) employed three different classifiers with a maximum uncertainty linear discriminant analysis, centroid, and average linkage methods. The experiments combined three different imaging features (image intensities, GM densities, and local volume changes), and the best accuracy achieved was 81.6%, showing an insignificant improvement compared to a single classifier accuracy of 80.6%.

Ensemble learning has also been used for the detection of other neurological diseases. Liu, Zhang, and Shen (2012) proposed a local patch-based subspace ensemble method to diagnose Alzheimer's disease and improved the accuracy by 3% versus a single classifier. Lebedev et al. (2014) reached 91% with random forests for the same disease, and Liu, Shang, Zheng, and Wen (2016) combined linear regression, linear SVM, naive Bayes, positron emission tomography (PET), and MRI for dementia diagnosis. They found 96.7% specificity for AD versus HC and more than 60% specificity for mild cognitive impairment versus HC. Gould et al. (2014) used resampling and created an ensemble of SVMs to ensure result stability while classifying the cognitive subtypes of schizophrenia.

Here, we use an ensemble learning method to improve the classification accuracy of an ANN, particularly a multilayer perceptron (MLP). Although there are many types of ensemble learning techniques in use, we adapt the random subspace ensemble (RSE) method (Ho, 1998) due to its expected applicability to the problem known as the curse of dimensionality in brain image classification (Lemm, Blankertz, Dickhaus, & Müller, 2011). Furthermore, we compare random subspace ensemble neural networks with SVM, one of the most common approaches used in pattern recognition. We refer to the combination of these methods as RSE-MLP and RSE-SVM in the following text.

The letter is organized into five sections. Section 2 describes the mathematical background—brain image preprocessing and the design of the ensemble learning classification framework including feature selection, classification and validation. Section 3 shows the results, section 4 discusses the outcomes, and section 5 concludes.

This section presents the data used for the experiment and summarizes the mathematical issues, which are important for the classification pipeline presented here.

The MR images were collected in the University Hospital Brno. Patients were interviewed in compliance with the International Statistical Classification of Disease and Related Health Problems (ICD-10), and subsequently their blood and urine samples were collected for toxicological, hematological, and biochemical testing. Subjects with abnormal findings were excluded from the data set. None of the subjects had a family or personal history of axis I psychiatric conditions. All subjects signed the informed consent, and the study was approved by the ethics committee (Janousova et al., 2016).

The data set contained 104 (52 SZ $+$ 52 healthy controls (HC)) T1-weighted images of the entire head obtained with a 1.5 T MR device and the following parameters: sagittal tomographic plane thickness of 1.17 mm, the in-plane resolution was 0.48 mm $×$ 0.48 mm, a 3D field of view contained 160 $×$ 512 $×$ 512 voxels, inversion recovery/gradient recalled (IR/GR) sequence, repetition time (TR) was 1700 ms, echo time (TE) was 3.93 ms, inversion time was 1100 ms, flip angle was 15$∘$⁠, and the field of view (FOV) was 246 $×$ 246 mm. The data were matched for age median (min-max): SZ 22.9 (17–40), HC 23.0 (18.2–37.8), and sex because all subjects were men. More details about the data set and image acquisition can be found in Janousova et al. (2016).

The GM tissue segments were obtained from all images following bias field inhomogeneity correction, spatial normalization, and segmentation (Ashburner & Friston, 2005) with the use of the VBM8 toolbox (http://dbm.neuro.uni-jena.de/vbm/) implemented in the SPM8 software package. Spatial normalization steps involved affine registration to standard statistical parametric mapping (SPM) T1 template followed by the fast diffeomorphic registration algorithm DARTEL. The gray matter (GM) tissue segments were modulated with Jacobian determinants calculated from the obtained spatial transformations to account for registration-related changes in local volumes. The modulated GM segment images were subsequently smoothed with an 8 mm full width at half maximum (FWHM) gaussian kernel to enable intersubject comparisons and to make the data distribution more normal.

Spatial normalization steps in DBM included the same affine registration algorithm as in VBM. After transforming all bias-corrected images into the stereotaxic space, our original high-dimensional deformable registration technique (Schwarz, Kasparek, Provaznik, & Jarkovsky, 2007) was used to compute the vector displacement fields that maximized the normalized mutual information between the images and the high-resolution single-subject template retrieved from the database of International Consortium for Brain Mapping (ICBM). The registration algorithm calculated local forces in each voxel and their regularization via a modified Rogelj's elastic-incremental spatial deformation model (for more details, see Schwarz et al., 2007). The resulting 3D displacement vector fields were converted to scalar fields by computing Jacobian determinants in each voxel of the stereotaxic space. After logarithmic transformation, the resulting positive and negative values refer to local volume changes caused by the deformation (i.e., expansions and contractions, respectively), whereas no deformation effect is observed where the values are close to zero.

Examples of the gray matter tissue densities (GM) and the local volume changes (DEF) in a 2D slice are shown with MR intensities of a normalized image (INT) in Figure 1.

Support vector machines serve here as a reference classifier. The configuration includes the maximum number of iterations, which was increased from the default 15,000 to 100,000 to ensure convergence of each SVM instance; linear kernel; constraint $C=1$⁠; and the sequential minimal optimization algorithm for training. The optimization of the $C$ parameter could potentially affect the results, but this was not considered for the same reason given in the previous paragraph.

The goals of this study include understanding (1) how the number of classifiers in an ensemble, size of a feature pool (FP), and length of feature input vector improve classification performance measures—overall accuracy (OA), sensitivity (SEN), and specificity (SPE); (2) which morphometry method, VBM or DBM, better suits the proposed classification scheme; and (3) if the combination of both morphometry methods (VBM and DBM) improves the classification success rate. Despite the differing nature of these morphometry methods, they seem to be complementary, and the combination of them could yield more variable information, making better conditions for finding complex discriminative information for adaptation of ensemble-based classifiers.

The experimental design is shown in Figure 2. The first part of the experiment involved random subspace ensembles combined with ANNs and SVM trained with the GM features extracted using the VBM pipeline. Details and results of this task were reported in Vyškovský et al. (2016). Here, we employed the DBM pipeline to extract the DEF features and obtain information different from the GM features. The next logical step was to employ both morphometry methods concurrently. This combination allowed us to extract information from both morphometry methods and thus to feed the classifiers with more variable features to create a decision boundary.

Feature selection and model adaptation followed brain image preprocessing using morphometry methods. These two steps were included in a validation loop of leave-one-out cross-validation (LOO-CV). Here, one subject serves as a testing sample, and the others are used for training. This process is repeated $N$ times. The performance of the classifiers is then measured on the test set. This validation method is widely used in neuroimaging; however, it is recommended that a more robust validation approach should be used, involving images acquired from different medical centers and devices with the same parameter settings (Nieuwenhuis et al., 2012).

The feature selection step chooses only the features with high discrimination power and is determined by a selected discrimination criterion—here, a simple two-sample $t$-test is used. This univariate method basically reveals the voxels in which the two groups significantly differ. We do not draw a line between significant and insignificant features, but we sort the features in ascending order based on the $p$-value and use the most significant ones. This application justifies the use of a parametric test, which is computationally faster than the Mann-Whitney U test. Since the images are smoothed during the VBM and DBM morphometry pipelines, autocorrelation among neighboring voxels may appear. However, the ability of single voxels to discriminate the patients from healthy controls is neither searched for nor interpreted (unlike, for example, logistic regression model coefficients in many applications). In addition, simplicity of the classification algorithm is preferred. The voxels are therefore not checked for autocorrelation.

After feature extraction and selection, a feature pool is prepared and classifiers are adapted. This is also where the ensemble learning is held. The algorithm of these processes has several steps:

1. The feature pool is created from the specified number of the most significant features. In this study, the influence of the size of FP on the performance measures was explored. The defined sizes were 10,000 and 100,000 voxels. Another way to define a feature pool size could be to subject all the significant voxels selected to the two-sample $t$-test applying FDR correction. This approach is not feasible here as it yields different numbers of voxels in each fold of cross-validation. The numbers of significant voxels (level of significance = 0.05) in our data set are min $=$ 194, max $=$ 10,320, median $=$ 1537.5 for VBM and min $=$ 0, max $=$ 86, median $=$ 0 for DBM. Therefore, a sufficient number of input features for classifiers cannot be guaranteed.

2. This FP played the role of a bag from which a feature vector (FV) was randomly chosen. Since the length of an FV directly affects the results, the three options of this parameter were defined and explored: 100, 1000, and 10,000.

3. The feature vector of each training subject and the corresponding label of the group were applied in the adaptation of the classification algorithms. This particular vector was designed to train both MLP and SVM to ensure the comparativeness of the results. After the adaptation phase, the testing subject was classified, and its predicted label was stored. Due to the random initialization of weights in MLP, this adaptation was performed 11 times, and the predicted class was based on voting. Thus, this step also included ensemble learning in which the variability was gained by random weight initialization.

4. Steps 2 and 3 were repeated 31 times, and thus many predictions based on randomly chosen feature vectors from FP were gained and could be used to vote on the final class for the testing subject. The number of voting classifiers (31) proved sufficient to demonstrate the overall trend toward greater accuracy when ensembles of classifiers were used to vote for the final class.

During these four steps, the results of single classifiers trained on different features were gained. In the next step, these results were used for voting. Since 31 single classifications were computed, any odd number of these results between 1 and 31 could be used for voting. Also, any combination could vote. Therefore, in order to generate smooth trend curves, all the trained classifiers were used for evaluation. For instance, there is only one ensemble of size 31, but there are 4495 ensembles containing three classifiers. All possible combinations, but not more than 10,000 (this limitation was added to save computational time), were used to compute the performance measures, and the outcomes were averaged. Furthermore, the experiment was performed 10 times because the feature vectors were chosen randomly from the FP and the results were averaged.

In this letter, we experimented with VBM and DBM morphometry methods, individually and in combination, adding other parameters, such as the size of the feature pool, the length of the feature vector, and the type of classifier. We studied the influence of these parameters on performance measures. The results are presented here.

The experiment was performed in Matlab R2015b on a computer with 2$×$ Intel Xeon CPU E5-2640 2.5 GHz and RAM 64 GB. To train one iteration of a framework, both MLP and SVM included, took around 20 hours (evaluation not included).

The first experiment was carried out on the features extracted from MRI images using VBM. A feature pool of size 10,000 (see Figure 3a) reached a maximal overall accuracy around 64.04 $±$ 0.81% (MLP, SEN $=$ 60.00%, SPE $=$ 68.08%). The use of shorter feature vectors revealed better outcomes compared to the use of the longest one for training the same classifier except for the SVM with 1000 inputs; the accuracy fell below SVM with 10,000 inputs. The ANN performed better than SVM when compared to models with the same number of input features. When the FP size was increased to 100,000 (see Figure 3b), the accuracy was increased in all models and reached 68.20 $±$ 0.24% (MLP, SEN $=$ 67.09%, SPE $=$ 69.32%). This time, models adapted on longer FV were better than those adapted on the shortest one. The neural network was a better model than SVM; however, SVM with 100 inputs improved the accuracy when more single models voted in the ensemble. With 19 voters, it outperformed the MLP. In both FP sizes, the RSE significantly helped with the classification accuracy only when short vectors were used.

In the second part of the experiment, we focused on the images preprocessed by DBM (see Figure 4). Here, the parameter settings are comparable to those in Figure 3. Regarding the FP of size 10,000 (see Figure 4a), the MLP in all cases was the better classifier when the same features were used (max. OA $=$ 66.46 $±$ 0.70%, SEN $=$ 64.88%, SPE $=$ 68.04%). The shorter the feature vector used to train the models, the higher the accuracy achieved except for the SVM trained on FV of the middle size. This was better than the SVM trained on the longest FV. The bigger feature pool (see Figure 4b) helped to increase the OA to 70.10 $±$ 0.96% (SEN $=$70.58%, SPE $=$ 69.62%). However, this time, the SVM yielded better results than MLP. The longer FV helped to gain better results than the shortest FV. A random subspace ensemble method helped only in the case of the shorter FV in both Figures 4a and 4b. Nevertheless, this improvement did not outperform models adapted on the longest FV in Figure 4b. The features extracted using DBM provided better information for the classification task than VBM because the accuracy increased to 70%.

The last experiment took advantage of both morphometry methods. Feature pools created from VBM and DBM were combined with 50% of the feature vector selected from FP based on VBM; the other 50% were selected from FP based on DBM. Thus, the sizes of FV were equal to previous experiments, which makes the models comparable (see Figure 5). In Figure 5a, the better classifier is SVM, especially for longer FV. The shortest FV in the MLP is better, but the gap between ANN and SVM decreases with the increasing number of voting models. This time, the classifiers based on longer FV reached OA of 72.88 $±$ 0.88% (SVM, SEN $=$ 67.69%, SPE $=$ 78.08%). In Figure 5b, both classifiers revealed similar results. Increasing the number of voters seemed to help only when the short FV were used, but this did not help to outperform classifiers based on longer FV. The combination of VBM and DBM improved the accuracy to 73.12 $±$ 0.16% (MLP, SEN $=$ 71.14%, SPE $=$ 75.10%) and 73.51 $±$ 0.76% (SVM, SEN $=$ 72.97%, SPE $=$ 74.05%).

The impact of the morphometry methods on classification accuracy was statistically assessed as follows. A number of classifiers were trained on the features using VBM, DBM, and the combination of both morphometry methods. The classifier that outperformed the other classifiers was chosen. Following that, a two-sample $t$-test was run to determine whether the difference in accuracy is statistically significant at the 0.05 alpha level. The results presented in Table 1 demonstrate that the DBM feature extraction method is better than VBM in terms of classification accuracy. What is more, the combination of both morphometry methods showed even better results, outperforming either method alone.

In this section, the feature pools were analyzed and visualized in this way. The FPs were computed for each fold of leave-one-out cross-validation separately, because even one subject separated during LOO-CV changes the shape of the feature pool. The frequency of voxel occurrence in a feature pool is visualized in 2D axial slices in Figure 6, demonstrating which parts of the brain were involved in training the classifiers.

Regarding VBM, the FPs overlapped mainly with these structures: frontal lobe (inferior frontal gyrus, medial frontal gyrus, superior frontal gyrus, middle frontal gyrus, and precentral gyrus), temporal lobe (superior temporal gyrus, and middle temporal gyrus), limbic lobe (cingulate gyrus), occipital lobe, and parietal lobe (mainly in 100,000 FP). In the case of DBM, the FPs overlapped with the following structures: frontal lobe, occipital lobe (cuneus, lingual gyrus, and middle Occipital Gyrus), temporal lobe (mainly in 100,000 FP), parietal lobe (mainly in 100,000 FP), cerebellum anterior lobe, culmen, limbic lobe (anterior cingulate), pons, temporal lobe (mainly in 100,000 FP), and parietal lobe (mainly in 100,000 FP).

As shown in Figure 6, the voxels were selected from multiple structures using both morphometry methods, even when the 10,000 most significant voxels were selected. VBM selected voxels from outer parts of the brain, while DBM did not. The differences identified between the feature extraction processes of the studied morphometry methods may be attributed to the fact that, DBM, unlike VBM, takes the whole brain approach and is thus able to detect abnormalities not only in gray matter but also in all brain compartments at once. This fact may also contribute to better accuracy achieved when a combination of both morphometry methods is used to extract the imaging features.

This letter extends our previous study (Vyškovský et al., 2016), where RSE-MLPs were applied in first-episode schizophrenia recognition in brain images preprocessed by VBM. Here, two novel contributions are presented: (1) RSE-MLP and RSE-SVM methods are employed to detect schizophrenia in MRI data preprocessed using DBM and (2) both VBM and DBM preprocessing methods are combined to provide features for subsequent classification. As expected, the resulting classification accuracy was influenced by all parameters: type of morphometry, size of feature pool, length of feature vector, and type of classifier.

Classifiers adapted on short feature vectors were similarly successful for both FP sizes. Furthermore, they were improved along with increasing ensemble size, especially when a few voters were added. The bigger size of the ensembles did not help that much. This was not the case of the classifiers using longer FV (1000, 10,000) for their adaptation. Their performance was improved negligibly using RSE-MLP and RSE-SVM. However, they performed much better when they had bigger FP available. These trends may be explained by the fact that a bigger feature pool provides uncorrelated features from many different parts of the brain, which may offer variable discriminative information. The classifiers needed to use more of these vectors to converge because the shorter FV were constituted of fewer voxels, and they are less likely to hold sufficient information from different parts of the brain. This is consistent with the literature (Liu et al., 2012) classifying Alzheimer's disease with faster convergence with longer FV. The multilayer perceptron seems to be better than SVM when short FV and VBM are used for adaptation.

Besides the dimension of the feature vector, feature pool, and type of classifier, the classification accuracy was influenced by methods for feature extraction. While models created on the basis of VBM reached an overall accuracy of 68.20%, models that took information from DBM reached 70.10%. Although the difference in accuracy was only around 2%, this suggested that there was variability in the information provided by these two methods. When both methods were combined, the accuracy improved to 73.12% (MLP) and 73.51% (SVM). As expected, the combination of morphometry methods added features with positive influence on the discrimination of patients in the first episode of schizophrenia (FES) and HC.

The best results were achieved by MLP (FP $=$ 100,000; length of FV, 10,000; number of MLPs in ensemble, 21); OA $=$ 73.12%; SEN $=$ 71.14%, SPE $=$ 75.10%, and SVM (FP $=$ 100,000; length of FV, 1000; number of SVMs in ensemble, 9); OA $=$ 73.51%, SEN $=$ 72.97%, and SPE $=$ 74.05%. However, similar accuracies were achieved using single MLP (OA $=$ 73.09%, SEN $=$ 71.10%, SPE $=$ 75.07%) and single SVM (OA $=$ 73.43%, SEN $=$ 73.86%, SPE $=$ 73.00%), both with 10,000 input features. Thus, the RSE did not help in an important way. The performance measures reached in this study are consistent with the results reported in (Ashburner et al. (1998), Bleich-Cohen et al. (2014), and Dluhoš et al. (2014).

A statistical comparison test done on the best-performing ensemble classifiers trained on the features extracted using two different morphometry methods and their combination revealed that the set of imaging features obtained using the combination of VBM and DBM yielded a statistically significant better overall accuracy than the features obtained by either morphometry method alone. This suggests that either morphometry method provides different information that can be used to distinguish between the schizophrenic patients and healthy controls.

Furthermore, DBM provides features with higher discrimination capabilities than VBM. The distinction in feature quality stems from the differing nature of these data extraction techniques. While DBM takes all brain tissue types into account, VBM can extract features only from separated gray matter, white matter, or cerebrospinal fluid. The findings might have important implications for further research employing morphometry methods in schizophrenia diagnostics.

Compared to other studies mentioned in section 1, our method achieved similar accuracy to Dluhoš et al. (2014), and Jafri and Calhoun (2006). Some other papers (Bleich-Cohen et al., 2014; Nieuwenhuis et al., 2012; Charpentier & Savio, 2010; Yang et al., 2010) achieved better results but suffered from small sample sizes (20–53 subjects), which can lead to high accuracy in some parameter settings (Nieuwenhuis et al., 2012; Schnack & Kahn, 2016). Studies based on the same data set as ours reported similar outcomes. Dluhoš et al. (2017) achieved accuracy between 65% (gray matter densities as features) and 70% (Jacobian determinants as features), though they improved the accuracy to 76% when weights of SVM models were averaged from multiple models adapted on data sets obtained from three different clinical sites. Unfortunately, we did not have a multisite data set available. Janousova et al. (2016) selected features using penalized linear discriminant analysis with resampling and obtained classification accuracy around 66% when adhering to the correct procedure of cross-validation. In a different study, Janousova et al. (2015), the achieved accuracy exceeded 80%. Besides gray matter densities and local deformations, they used MR intensities, and the features were extracted using intersubject principal component analysis. The results of this study suggest that adding further feature extraction methods to our framework may improve the outcomes.

Although good classification accuracy was reached, confounding factors and limitations must be taken into consideration. The possible confounders include age, sex, long-term antipsychotic treatment, and disease progression. These were excluded because the patients were age matched, all male, and in FES. The first limitation is related to the sample size. Nieuwenhuis et al. (2012) experimented with bootstrapping of samples and recommended at least 130 samples to gain robust and stable outcomes. The second limitation regarding the sample is that the validation was not done on the images acquired from different medical centers and devices with the same parameter settings, which would be considered a more appropriate validation method (Nieuwenhuis et al., 2012). Third, the classifiers used in the proposed framework depend on several parameters. The performance of SVM is influenced by C-parameter (Franke, Ziegler, Klöppel, & Gaser, 2010), and MLP's success depends on the number of neurons, layers, types of activation functions, and other training parameters. All of these parameters were fixed here because the computational cost to investigate them would have risen markedly otherwise.

The success of RSE-based classification of FES is dependent on several crucial steps, including brain preprocessing and feature extraction, size of the feature pool, length of the feature vector, and type of classifier. In this study, VBM and DBM feature extraction methods and their combination were employed to provide data for learning. As expected, their combination enabled the classifiers to capture more complex discriminative information compared to separate variants. Other information, for example, from psychological tests, could further improve the outcomes. In all the presented experiments, random subspace ensemble-based approaches helped only when the feature vectors were short. More suitable classifiers among MLP and SVM were not found, their success depended on other parameters in the proposed design.

This work was supported by research grant 17-33136A from the Ministry of Health, Czech Republic.