fMRI classification method with multiple feature fusion based on minimum spanning tree analysis

https://doi.org/10.1016/j.pscychresns.2018.05.001Get rights and content

Highlights

  • The MSTs of MDD patients were more like random networks than NC.

  • Compared with NCs, the MSTs of MDD patients exhibited significant differences in certain regions concentrated in the LCSPT circuit.

  • The results of prosed method exhibited significantly better performance than the methods using a single feature type.

  • The proposed method performed better than the conventional methods of constructing the functional network by partial correlations or Pearson correlations.

Abstract

Resting state functional brain networks have been widely studied in brain disease research. Conventional network analysis methods are hampered by differences in network size, density and normalization. Minimum spanning tree (MST) analysis has been recently suggested to ameliorate these limitations. Moreover, common MST analysis methods involve calculating quantifiable attributes and selecting these attributes as features in the classification. However, a disadvantage of these methods is that information about the topology of the network is not fully considered, limiting further improvement of classification performance. To address this issue, we propose a novel method combining brain region and subgraph features for classification, utilizing two feature types to quantify two properties of the network. We experimentally validated our proposed method using a major depressive disorder (MDD) patient dataset. The results indicated that MSTs of MDD patients were more similar to random networks and exhibited significant differences in certain regions involved in the limbic-cortical-striatal-pallidal-thalamic (LCSPT) circuit, which is considered to be a major pathological circuit of depression. Moreover, we demonstrated that this novel classification method could effectively improve classification accuracy and provide better interpretability. Overall, the current study demonstrated that different forms of feature representation provide complementary information.

Introduction

The human brain is a complex system with a sophisticated structure. As a non-invasive way to measure spontaneous neural activity in the human brain, resting-state functional magnetic resonance imaging has attracted considerable attention (Fox et al., 2007, Wang et al., 2010). Resting-state fMRI using blood oxygenation level dependent (BOLD) signals as neurophysiological indicators can detect spontaneous low-frequency brain activity and has been successfully applied to the diagnosis of neuropsychiatric diseases such as epilepsy (Horstmann et al., 2010, Raj et al., 2010), Alzheimer's disease (AD) (He et al., 2008, Stam, 2010, Supekar et al., 2008), schizophrenia (Lynall et al., 2010, Micheloyannis et al., 2006, Rubinov et al., 2009), and so on. Functional connectivity is defined as the “temporal correlations between spatially remote neurophysiological events” (Friston et al., 1993). Functional connectivity models can provide a cost-effective method for information differentiation and integration (Achard and Bullmore, 2007). Graph theory is a commonly used tool for analyzing neural network topology, providing a comprehensive and sophisticated framework for characterizing network topology. By applying graph theory, we can gain insight into normal brain architecture and pathological mechanisms for brain disorders and the ways that brain networks exhibit features of both integration and segregation of information processing (Bullmore and Sporns, 2012, Rubinov and Sporns, 2010, Stam and van Straaten, 2012, van Straaten and Stam, 2013). Several studies have reported that brain networks exhibit many important topological properties, such as “small world” and modular attributes (Achard and Bullmore, 2007, Chen et al., 2008, He et al., 2007, Salvador et al., 2005). Moreover, research into a range of brain diseases, including schizophrenia (Lynall et al., 2010), Alzheimer's disease (Supekar et al., 2008) and epilepsy (Douw et al., 2013, Edwin et al., 2014, Lee et al., 2006) has revealed abnormal changes in the network topology of patients.

Although conventional graph theory analysis is helpful for understanding disease mechanisms (Bullmore and Sporns, 2012), it is limited by standardization and methodological difficulties when comparing different groups or conditions (Fornito et al., 2013, van Wijk et al., 2010). For instance, graph measures are affected by the scale of the network (the number of nodes), network sparsity (the percentage of links) and the average degree (the number of connections per node). A commonly used approach to correct for size or average degree dependence is to normalize graph metrics by random graphs (van Wijk et al., 2010). However, this normalization method does not solve the dependence on size, degree or density effects, and may even exacerbate it (Tewarie et al., 2014). In addition, using fixed nodes and average degree in the network can eliminate size effects, but may introduce spurious connections or ignore strong connections in the network (van Wijk et al., 2010). Even the use of weighted graphs instead of unweighted graphs does not provide an optimal solution, since measures computed on these graphs are affected by the large number of noisy connections and the average functional connectivity strength (Tewarie et al., 2014, Tewarie et al., 2015).

To avoid the shortcomings of traditional graph theory, researchers have applied the minimum spanning tree (MST) method (Jackson and Read, 2010a, Jackson and Read, 2010b), utilizing a sub-graph of the original network to analyze brain networks. An MST is mathematically defined as an acyclic subgraph that connects all nodes, with minimal link weights (Kruskal, 1956, Prim, 2010). The MST method, which avoids methodological biases, is particularly suitable for the comparison of brain networks (Tewarie et al., 2014). Importantly, the link weights in neuroimaging data typically represent connectivity strength, which can be considered an inverse distance. Because high edge weights represent strong temporal correlation between two brain regions, an MST based on functional connectivity strength is formally considered a maximum spanning tree with maximal link weights or the longest path between its two endpoints.

Test-retest reliability is an important concept in social, behavioral, physical, biological and medical sciences. Since various factors interfere with the actual measurement procedure, test-retest reliability is particularly important for building and choosing reliable measures (Nakagawa and Schielzeth, 2010). Test-retest reliability of the network metrics based on the fMRI functional networks is of significant value in the neuroscience community (Chen et al., 2015, Zuo et al., 2014, Zuo et al., 2012, Zuo and Xing, 2014). Many researchers world-wide presented connectome biomarkers for various brain disorders and in view of this, their reproducibility should be explored (Dimitriadis et al., 2016a, Dimitriadis et al., 2016b, Dimitriadis et al., 2015a, Dimitriadis et al., 2015b, Dimitriadis et al., 2015c). High test-retest reliability is crucial for the development of biomarker-based clinical tests for early detection, timely interventions and diagnoses of brain disorders, especially psychiatric disorders, which currently lack a gold standard biological definition (Kaiser, 2013). MST is a new and emerging method to human functional connectomics, which has been proved to be an unbiased method for brain networks in order to get reliable network metrics (Tewarie et al., 2014). In particularly, it has been recently demonstrated that such a method can be highly test-retest reliable (Dimitriadis et al., 2017, Tewarie et al., 2015).

To our knowledge, Lee and colleagues (Lee et al., 2006) were the first group to use MST as a tool for analyzing complex functional brain networks on the basis of electroencephalography (EEG) data. To date, the MST method has been used in several recent studies on development and neuropsychiatric diseases (Boersma et al., 2013, Dubbelink et al., 2013, Edwin et al., 2014, Ortega et al., 2008, Stam, 2014, Tewarie et al., 2013). For example, MST network analyses have been used to identify the critical nodes in a temporal network associated with seizures (Ortega et al., 2008) and to characterize the different network topology in different epilepsy types (Lee et al., 2006). In addition, constructing MST based on an original graph has enabled researchers to capture default mode network changes in Alzheimer's disease (Çiftçi, 2011). A recent study of functional network changes during brain maturation in children revealed that MST network analyses were sensitive for detecting changes in network topology, producing results that were consistent with conventional graph theoretical measurements of the same data (Boersma et al., 2013).

Common methods of MST analysis involve the calculation of quantifiable attributes, such as local metrics (including degree, betweenness centrality and eccentricity) and global metrics (including diameter, leaf fraction, hierarchy and average eccentricity), and analysis of these attributes to detect significant between-group differences. A method based on brain region features is then used for classification. Network-based classification approaches based on brain region features have been widely studied in brain disease research. These methods involve two main steps: (1) extracting meaningful features from the connectivity network, followed by feature selection to select the most discriminative feature subset; and (2) training a classifier using the chosen features. However, the extracted features, such as local network measures and weights between region-of-interest (ROI) pairs, are usually concatenated into a long feature vector for subsequent feature selection and classification, without considering the disease-related topological information such as local and global topological information of the network (Chen et al., 2011, Wee et al., 2012). This may deteriorate the final classification performance (Chen et al., 2011, Jie et al., 2014b, Wee et al., 2012). In addition, it is difficult to find abnormal subnetworks in a whole network due to the complexity of brain networks. To overcome this limitation, we used a discriminative subnetwork of MST in this study. Discriminative subnetworks can effectively identify disorder patterns crossing several nodes (regions) in networks, as reported in many previous network classification studies, including studies of chemical compounds (Deshpande et al., 2003),and protein-protein interaction (PPI) networks (Zou et al., 2010). Research methods based on subgraphs do not need to be limited to a single brain region in the description of the brain network, meaning that they not only retain the topological information of the original sample, but do not lose the original discriminant information.

Importantly, there is a loss of sample information during classification using methods based on brain region features, as well as in the subgraph-based method. A specific limitation of methods based on brain region features is that they are missing topological information of the whole brain network. Meanwhile, subgraph-based methods are insensitive to changes in single brain regions. Because the brain network is a complex structure, the biological characteristics of brain networks cannot be fully acquired by applying a single feature (Sporns, 2011).

To overcome these limitations, in the current paper we propose a novel method using both brain region and subgraph features for classifying MDD patients from a healthy control group, using MST analysis. Here, we computed degree, eccentricity, and betweenness centrality of each brain region as local network features and extracted the subgraph features from MST networks. Then, we combined the two types of features, and chose a multi-kernel support vector machine (SVM) as the classifier to achieve better classification performance. The experimental results confirmed the feasibility of the proposed MST-based method. In addition, the classification results showed that our proposed method achieved better classification performance than a method using only a single feature type, and had better interpretability.

Section snippets

Participants

Subjects in our study included 38 first-episode drug-naive MDD patients (15 males and 23 females), and 28 healthy controls (13 males and 15 females) and were gender-, education- and age-matched (see Table 1). All patients were inpatients at the Department of Psychiatry, First Hospital of Shanxi Medical University, clinically diagnosed with MDD by at least two experienced psychiatrists according to Diagnostic and Statistical Manual of Mental Disorders, 4th ed. (First et al., 1997). The severity

Global metrics

By analyzing four global attributes of MST network, we were able to examine network topology differences between the control and patient groups. The results showed that the diameter and mean eccentricity values of MST were significantly lower in MDD patients compared with normal controls. However, leaf fraction and tree hierarchy values were significantly higher in MDD patients (P < 0.01, FDR corrected; Fig. 2).

Nodal metrics

Compared with normal controls, MDD patients exhibited significantly increased

Global metrics analysis

By applying the MST method, we were able to identify functional brain network differences between MDD patients and healthy controls. Analysis of four global attributes of the MST network revealed a clear difference between the control group and the MDD patient group. Specifically, MSTs of MDD patients exhibited a smaller diameter, reduced eccentricity, greater leaf number and greater tree hierarchy (Fig. 2). This suggests that the MST topology of the MDD group was more star-like, corresponding

Conclusion

In summary, the current results demonstrated the feasibility of our proposed multiple feature fusion classification method based on MST analysis, when evaluated with a MDD patient dataset. The experimental results revealed that our proposed method exhibited significantly better classification performance that methods using a single feature type. We constructed brain networks using an MST method that avoids normalization problems and methodological biases. Furthermore, applying a multiple

Author contributions

The sponsors had no role in the design or execution of the study, the collection, management, analysis, and interpretation of the data, or preparation, review, and approval of the manuscript. HG was responsible for the study design and writing the manuscript. PY performed data analysis and statistical processing. YX provided and integrated experimental data. JC supervised the paper. JX was the heads of the funds and supervised the paper. All authors approved the final version of the manuscript.

Disclosures

This manuscript has not been published or presented elsewhere in part or in entirety, and is not under consideration by any another journal. This study was carried out in accordance with the recommendations of the medical ethics committee of Shanxi Province (reference number: 2012013). All subjects have been given written informed consent in accordance with the Declaration of Helsinki. Meanwhile, all the authors have read through the manuscript, approved it for publication, and declared no

Acknowledgements

Financial disclosure: This study was supported by research grants from the National Natural Science Foundation of China (61373101, 61472270, 61402318, 61672374, 61741212), Natural Science Foundation of Shanxi Province (201601D021073) and the Scientific and Technological Innovation Programs of Higher Education Institutions in Shanxi (2016139).

References (116)

  • K.T. Kronmüller et al.

    Hippocampal volume in first episode and recurrent depression

    Psychiatry Res.

    (2009)
  • A.L. Lacerda et al.

    Anatomic evaluation of the orbitofrontal cortex in major depressive disorder

    Biol. Psychiatry

    (2004)
  • T. Lai et al.

    Reduction of orbital frontal cortex volume in geriatric depression

    Biol. Psychiatry

    (2000)
  • L. Li et al.

    Prefrontal white matter abnormalities in young adult with major depressive disorder: a diffusion tensor imaging study

    Brain Res.

    (2007)
  • M. Liu et al.

    Multitemplate-based multiview learning for Alzheimer's disease diagnosis

    Mach. Learn. Med. Imaging

    (2016)
  • S. Micheloyannis et al.

    Small-world networks and disturbed functional connectivity in schizophrenia

    Schizophr. Res.

    (2006)
  • K. Murphy et al.

    The impact of global signal regression on resting state correlations: are anti-correlated networks introduced?

    Neuroimage

    (2009)
  • G.J. Ortega et al.

    Complex network analysis of human ECoG data

    Neurosci. Lett.

    (2008)
  • A. Raj et al.

    Network-level analysis of cortical thickness of the epileptic brain

    Neuroimage

    (2010)
  • M. Rubinov et al.

    Complex network measures of brain connectivity: uses and interpretations

    Neuroimage

    (2010)
  • Y.I. Sheline

    Neuroimaging studies of mood disorder effects on the brain

    Biol. Psychiatry

    (2003)
  • C.J. Stam

    Use of magnetoencephalography (MEG) to study functional brain networks in neurodegenerative disorders

    J. Neurol. Sci.

    (2010)
  • C.J. Stam et al.

    The trees and the forest: Characterization of complex brain networks with minimum spanning trees

    Int. J. Psychophysiol.

    (2014)
  • C.J. Stam et al.

    The organization of physiological brain networks

    Clin. Neurophysiol.

    (2012)
  • J.D. Steele et al.

    Prefrontal cortical functional abnormality in major depressive disorder: a stereotactic meta-analysis

    J. Affect. Disord.

    (2007)
  • P. Tewarie et al.

    Functional brain network analysis using minimum spanning trees in Multiple Sclerosis: an MEG source-space study

    Neuroimage

    (2014)
  • P. Tewarie et al.

    The minimum spanning tree: an unbiased method for brain network analysis

    Neuroimage

    (2015)
  • N. Tzourio-Mazoyer et al.

    Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain

    Neuroimage

    (2002)
  • E.C. van Straaten et al.

    Structure out of chaos: functional brain network analysis with EEG, MEG, and functional MRI

    Eur. Neuropsychopharmacol.

    (2013)
  • O. Abe et al.

    Voxel-based analyses of gray/white matter volume and diffusion tensor data in major depression

    Psychiatry Res.

    (2009)
  • S. Achard et al.

    Efficiency and cost of economical brain functional networks

    PLoS Comput. Biol.

    (2007)
  • G.S. Alexopoulos et al.

    Research advances in geriatric depression

    World Psychiatry

    (2009)
  • M.A. Alvarez et al.

    A shortest-path graph kernel for estimating gene product semantic similarity

    J. Biomed. Semant.

    (2011)
  • Y. Benjamini et al.

    Controlling the false discovery rate: A practical and powerful approach to multiple testing

    J. R. Sta. Soc.

    (1995)
  • R. Bluhm et al.

    Resting state default-mode network connectivity in early depression using a seed region-of-interest analysis: decreased connectivity with caudate nucleus

    Psychiatry Clin. Neurosci.

    (2009)
  • M. Boersma et al.

    Growing trees in child brains: graph theoretical analysis of electroencephalography-derived minimum spanning tree in 5- and 7-year-old children reflects brain maturation

    Brain Connect

    (2013)
  • M. Boersma et al.

    Network analysis of resting state EEG in the developing young brain: structure comes with maturation

    Hum. Brain Mapp.

    (2011)
  • V. Borchardt et al.

    Preprocessing strategy influences graph-based exploration of altered functional networks in MDD

  • K.M. Borgwardt et al.

    Protein function prediction via graph kernels

    Bioinformatics

    (2005)
  • E. Bullmore et al.

    The economy of brain network organization

    Nat. Rev. Neurosci.

    (2012)
  • C.J.C. Burges

    A tutorial on support vector machines for pattern recognition

    Data Mining Knowl. Discovery

    (1998)
  • B. Chen et al.

    Individual variability and test-retest reliability revealed by ten repeated resting-state brain scans over one month

    PLoS One

    (2015)
  • G. Chen et al.

    Classification of Alzheimer disease, mild cognitive impairment, and normal cognitive status with large-scale network analysis based on resting-state functional MR imaging

    Radiology

    (2011)
  • Z.J. Chen et al.

    Revealing modular architecture of human brain structural networks by using cortical thickness from MRI

    Cereb. Cortex

    (2008)
  • K. Çiftçi

    Minimum spanning tree reflects the alterations of the default mode network during Alzheimer's disease

    Ann. Biomed. Eng.

    (2011)
  • W. Coryell et al.

    Subgenual prefrontal cortex volumes in major depressive disorder and schizophrenia: diagnostic specificity and prognostic implications

    Am. J. Psychiatry

    (2005)
  • R.J. Davidson et al.

    The neural substrates of affective processing in depressed patients treated with venlafaxine

    Am. J. Psychiatry

    (2003)
  • M. Demuru et al.

    Brain network analysis of EEG functional connectivity during imagery hand movements

    J. Integr. Neurosci.

    (2013)
  • M. Deshpande et al.

    Frequent sub-structure-based approaches for classifying chemical compounds

  • S. Dimitriadis et al.

    Revealing cross-frequency causal interactions during a mental arithmetic task through symbolic transfer entropy: a novel vector-quantization approach

    IEEE Trans. Neural Syst. Rehabil. Eng.

    (2016)
  • Cited by (0)

    View full text