Identifying differences in brain activities and an accurate detection of autism spectrum disorder using resting state functional-magnetic resonance imaging : A spatial filtering approach
Graphical abstract
Introduction
Autism Spectrum Disorders (ASD) represent a cluster of relatively common developmental disorders characterized by impaired social communication/social reciprocity and restricted/repetitive stereotyped behavioral patterns. ASD is one of the major problems affecting children and it has been shown recently that nearly 1 in 68 children are affected (CDC, 2014). Traditionally, these disorders are diagnosed using interview based methods such as the Autism Diagnostic Observation Schedule (Lord, 1989) and the Autism Diagnostic Interview - Revised (Lord et al., 1994). However, these methods are error prone and they are also unable to point out any biological basis behind the observed behavioral symptoms. Knowledge of such biological factors is useful for treatment, therapy or possibly even for prevention. To overcome these difficulties, brain imaging based methods are being explored as alternate diagnostic tools (Menon, 2011, Mostapha, Casanova, Gimel’farb, El-Baz, 2015), to automatically detect ASD in a non-invasive fashion and also identify the possible biological factors behind these conditions.
Magnetic Resonance Imaging (MRI) is an important brain imaging method that offers high resolution information about the structure, composition and functioning of the brain. Recently, functional MRI (fMRI) of the brain obtained during the resting state has been explored to gain insights into the neural activities behind many neuro-developmental disorders including ASD (Menon, 2011). In this approach, by using fMRI, the various neural connections/pathways are identified by measuring the temporal correlations in the Blood Oxygen Level Dependent (BOLD) signals from the different regions of the brain when the person is in a state of cognitive rest. The spontaneous fluctuations in the BOLD signals during resting state are considered as strong indicators of the organization and maintenance of functional brain systems (Fox, Raichle, 2007, Raichle, 2010) as well as information processing (Greicius and Menon, 2004). The idea that cortical under-connectivity might be underlying autism was initially studied using task-based fMRI studies (Just et al., 2004). Subsequent studies (Cherkassky, Kana, Keller, Just, 2006, Kennedy, Courchesne, 2008, Monk, Peltier, Wiggins, Weng, Carrasco, Risi, Lord, 2009) using the resting state fMRI also examined the effects of ASD on the functional connectivity of the brain and found significant differences that could be related to some of the core-symptoms observed in ASD patients.
For the analysis of connectivity from the BOLD time-series, a connectivity matrix is generally used to quantify the strengths of the functional connections between the various regions of the brain. The entries in the connectivity matrix are based on a statistical measure (Pearson correlation coefficient, mutual information, etc.) of the inter-regional activations of the brain during the fMRI acquisition. Several attempts have been made recently to study the connectivity of the brain under ASD as reported in Uddin et al. (2013b); Müller et al. (2011); Uddin et al. (2010). By using small sets of fMRI data (acquired from lower than 100 subjects), many of these works have attempted to explore the phenomena of weaker connections (hypo-connectivity) or stronger connections (hyper-connectivity) due to ASD between the different regions of the brain. While some of these studies indicate hypo-connectivity for ASD patients (Assaf, Jagannathan, Calhoun, Miller, Stevens, Sahl, O’Boyle, Schultz, Pearlson, 2010, Cardinale, Shih, Fishman, Ford, Müller, 2013, Mueller, Keeser, Samson, Kirsch, Blautzik, Grothe, Erat, Hegenloh, Coates, Reiser, Hennig-Fast, Meindl, 2013, Barttfeld, Wicker, Cukier, Navarta, Lew, Leiguarda, Sigman, 2012, Weng, Wiggins, Peltier, Carrasco, Risi, Lord, Monk, 2010), some other studies have also indicated hyper-connectivity (Di Martino, Kelly, Grzadzinski, Zuo, Mennes, Mairena, Lord, Castellanos, Milham, 2011, Uddin, Supekar, Lynch, Khouzam, Phillips, Feinstein, Ryali, Menon, 2013, Lynch, Uddin, Supekar, Khouzam, Phillips, Menon, 2013, Washington, Gordon, Brar, Warburton, Sawyer, Wolfe, Mease-Ference, Girton, Hailu, Mbwana, et al., 2014) in certain regions of the brain for ASD patients. Interestingly, Tyszka et al., reported no differences in functional connectivity at the whole brain level (Tyszka et al., 2014) while Gotts et al. reported whole brain hypo-connectivity (Gotts et al., 2012). While most studies have indicated differences in connectivity, it is not clear whether ASD patients suffer from hypoconnectivity or hyperconnectivity and the studies in this regard have been inconclusive so far. By extracting features that represent the differences in connectivity, some studies have also attempted to use machine learning methods to classify ASD patients from neurotypical subjects. In these studies (Anderson, Nielsen, Froehlich, DuBray, Druzgal, Cariello, Cooperrider, Zielinski, Ravichandran, Fletcher, et al., 2011, Barttfeld, Wicker, Cukier, Navarta, Lew, Leiguarda, Sigman, 2012, Murdaugh, Shinkareva, Deshpande, Wang, Pennick, Kana, 2011, Uddin, Supekar, Lynch, Khouzam, Phillips, Feinstein, Ryali, Menon, 2013, Dodero, Sambataro, Murino, Sona, 2015), ASD classification accuracies in the range of 78–91% have been reported. However, the use of small sets of proprietary data limits their capabilities to provide more generalized findings.
To enable a large scale investigation of ASD from fMRI, recently, a large dataset of resting state fMRI volumes obtained from different locations has been publicly released by the Autism Brain Imaging Data Exchange (ABIDE, 2013) consortium. Based on this dataset, Di Martino et al. have examined the whole-brain and regional connectivity differences due to ASD (Di Martino et al., 2014). To avoid the computational complexity involved in whole-brain voxel wise analysis, a functional and structural parcellation based approach was adopted in Di Martino et al. (2014) based on standard atlases (Kennedy, Lange, Makris, Bates, Meyer, Caviness, 1998, Craddock, James, Holtzheimer, Hu, Mayberg, 2012) of the human brain. This study provided evidence of wide-spread hypoconnectivity throughout the brain and also some confined regions of hyperconnectivity due to ASD.
Detailed ASD classification analysis using the fMRI volumes from the complete ABIDE dataset has been studied only in Nielsen et al. (2013). In this study, the pairwise correlation coefficients between 7266 gray matter voxels in the brain were used as features and a general linear model was used to achieve a classification accuracy of 60%. Since a voxel based approach was used in this study, the resultant feature vectors were found to be of high dimensionality, where many of the features may not have the relevant discriminative ability. To overcome this, in another recent study (Iidaka, 2015), the dimensionality of the feature space was reduced by considering only 90 regions in the brain, according to the Automated Anatomical Labeling (AAL) template (Tzourio-Mazoyer et al., 2002). An effect-size based feature selection approach further lowered the number of features and a classification accuracy of 87% was obtained. However, in this study (Iidaka, 2015), only the fMRI data from the adolescent subjects from ABIDE whose age falls below 20 years were considered. Other studies (Price, Wee, Gao, Shen, 2014, Chen, Keown, Jahedi, Nair, Pflieger, Bailey, Müller, 2015) based on even smaller subsets of data (<250 subjects) from ABIDE have also reported a classification accuracy around 90%.
In all the above studies, the pairwise correlation coefficients of the BOLD time-series signals obtained from the different regions of the brain were directly used for the analysis of connectivity and also used as the features to classify ASD patients from neurotypical subjects. This approach generally produces a large set of connections/features from which it is difficult to identify the discriminative brain activities underlying ASD. To obtain the discriminative brain activities, the BOLD time-series signals need to be projected into a new space such that the components representing the differences in underlying neural activities are clearly separated. Based on the projected signals, relevant discriminative and interpretable features may be extracted for an accurate classification. In this paper, such a method referred to as the Spatial Feature based detection Method (SFM) is developed, which adopts a spatial filtering approach to extract the most discriminative components of the BOLD time-series signals and also show clearly, the differences in brain activities.
In SFM, the mean connectivity matrices of the ASD patients and the neurotypical subjects are projected orthogonally such that maximum separability is achieved between the variances of the signals from the two classes. Such a projection matrix is obtained using the ‘Fukunaga–Koontz’ transform, which also has been widely used in Electro-Encephalogram (EEG) signal analysis (Koles, 1991, Koles, Lazar, Zhou, 1990, Ramoser, Muller-Gerking, Pfurtscheller, 2000). This projection matrix (also known as a spatial filter) is used to project the BOLD time-series signals and extract the temporal signal components that account for the most significant differences in the underlying neural activities between the ASD patients and neurotypical subjects. The inverse of the spatial filter provides the spatial distribution weights for the projected time-series and hence they point out the regions in the brain that are primarily responsible for the differences in the BOLD time-series signals. By interpolation, one can obtain a topographical map of the regions in the brain that are responsible for the functional differences due to ASD. Such a spatial pattern map also provides new insights on the variations in neural activities which cannot be found by just observing the pairwise correlation co-efficients alone, as reported by the previous studies. Highly discriminative features are then obtained from the log(variance) of the projected BOLD time-series signals and the number of features are further reduced by selecting only the top m signals from each class that account for the most significant differences in the variances with respect to the other class. The extracted features are then used to obtain an accurate classification of ASD using a Support Vector Machine (SVM) classifier.
Recent studies in the medical literature (Jacquemont, Coe, Hersch, Duyzend, Krumm, Bergmann, Beckmann, Rosenfeld, Eichler, 2014, Fair, Cohen, Dosenbach, Church, Miezin, Barch, Raichle, Petersen, Schlaggar, 2008, Supekar, Uddin, Prater, Amin, Greicius, Menon, 2010) have shown the gender and age-group specific effects of ASD. Hence, to obtain meaningful insights, in this paper, comprehensive studies have been performed by considering different categories of gender (males/females) and age-group (adolescents/adults) separately. The ABIDE dataset has been used for the performance evaluation of SFM. Specifically, the publicly available preprocessed data from the Preprocessed Connectomes Project (PCP) (PCP, 2014) is used. The spatial patterns from SFM are used to identify those regions in the brain that are mainly responsible for differences in the BOLD signals obtained from the resting state fMRI. Among the males, the results clearly indicate a shift in the activities to the prefrontal cortex for males with ASD while some other parts of the brain show diminished activities compared to neurotypical subjects. Among females, such a clear shift is not evident; however, several regions, especially in the posterior and medial portions of the brain show diminished activities due to ASD. It is also found that better classification performance is obtained for females (95% for adults and 86.7% adolescents) when compared to males (78.6% for adolescents and 85.4% for adults) and within each gender better classification performance is obtained for adults when compared to adolescents. Based on this study, one can conclude that SFM achieves a higher classification accuracy, using lower number of features compared to the earlier methods.
The paper is organized as follows: Section 2 provides the details of the preprocessed fMRI dataset and the principles behind the SFM approach. The procedure adopted by SFM to identify the distinguishable brain activities due to ASD and also extract discriminative features for accurate ASD detection are explained in detail. Using SFM, Section 3 gives a comprehensive analysis of the resting state fMRI data from ASD patients and neurotypical subjects from the ABIDE dataset. For each category of gender and age-group, the spatial pattern maps of those brain regions where significant functional differences have been identified due to ASD are first highlighted. The classification performance results using the discriminative features obtained from SFM are then provided. Finally, the conclusions from this study are summarized in Section 4.
Section snippets
Materials and methods
This section provides the details of the fMRI data used in this study for ASD detection and also the approaches used to process the raw data. Section 2.1 provides the details of the fMRI dataset used in this paper. From these fMRI images, the Blood Oxygen Level Dependent (BOLD) signals (time-series data) are generated. The details of the preprocessing steps to generate the BOLD signals are given in Section 2.1.1. Finally, Section 2.2 describes the Spatial Feature based detection Method (SFM)
Results and discussion
In this section, the SFM approach is used to analyze the resting state fMRI scans available in the ABIDE (PCP, 2014) dataset. Using the SFM approach, the regions in the brain that exhibit distinguishable brain activities during the resting state are first highlighted using the spatial pattern maps. Since several studies in the medical literature have pointed out the gender and age-group specific effects of ASD, studies have been performed here by considering separately, the four different
Conclusions
In this paper a new approach referred to as the Spatial Feature based detection Method (SFM) is presented for an accurate diagnosis of ASD from the fMRI images. Using the large scale ABIDE fMRI dataset, detailed studies have been conducted on the performance of SFM for ASD diagnosis by separately considering both gender and age-groups. The spatial pattern maps obtained from SFM, provide new insights into the regions where differences in connectivity between ASD patients and neurotypical
Acknowledgments
We thank the anonymous reviewers for their valuable comments which improved the paper. We also thank the ABIDE and PCP initiatives which provided access to preprocessed fMRI data. The ABIDE dataset results from a sharing initiative from 17 international sites, with anonymized data being hosted by the International Neuroimaging Data-sharing Initiative (INDI). Funding sources for each individual sites are provided on the official ABIDE website (http://fcon_1000.projects.nitrc.org/indi/abide/).
References (70)
A fast diffeomorphic image registration algorithm
Neuroimage
(2007)- et al.
Unified segmentation
Neuroimage
(2005) - et al.
Abnormal functional connectivity of default mode sub-networks in autism spectrum disorder patients
Neuroimage
(2010) - et al.
State-dependent changes of connectivity patterns and functional brain network topology in autism spectrum disorder
Neuropsychologia
(2012) - et al.
The neuro bureau preprocessing initiative: open sharing of preprocessed neuroimaging data and derivatives
Frontiers in Neuroinformatics (Neuroinformatics 2013)
(2013) - et al.
Kernel-based analysis of functional brain connectivity on grassmann manifold
Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015
(2015) - et al.
Sentence comprehension in autism: thinking in pictures with decreased functional connectivity
Brain
(2006) - et al.
Inhibitory control in high-functioning autism: decreased activation and underconnectivity in inhibition networks
Biol. Psychiatry
(2007) - et al.
Abnormal functional connectivity in autism spectrum disorders during face processing
Brain
(2008) Autism Diagnostic Observation Schedule: a standardised observation of communicative and social behaviour
J. Autism Dev. Disord.
(1989)
Default mode network in childhood autism: Posteromedial cortex heterogeneity and relationship with social deficits
Biol. Psychiatry
Convergent findings of altered functional and structural brain connectivity in individuals with high functioning autism: a multimodal MRI study
PLoS ONE
Underconnected, but how? A survey of functional connectivity MRI studies in autism spectrum disorders
Cereb. Cortex
Differential deactivation during mentalizing and classification of autism based on default mode network connectivity.
PloS One
Optimal spatial filtering of single trial EEG during imagined hand movement
IEEE Trans. Rehab. Eng.
Abnormal ventral temporal cortical activity during face discrimination among individuals with autism and asperger syndrome
Arch. Gen. Psychiatry
Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the MNI MRI single-subject brain
Neuroimage
Reconceptualizing functional brain connectivity in autism from a developmental perspective
Front. Hum. Neurosci.
From connectivity models to region labels: identifying foci of a neurological disorder
IEEE Trans. Med. Imaging
Neural mechanisms of imitation and ‘mirror neuron’ functioning in autistic spectrum disorder
Neuropsychologia
Decreased interhemispheric functional connectivity in autism
Cereb. Cortex
Functional connectivity magnetic resonance imaging classification of autism
Brain
Superior temporal gyrus, language function, and autism
Dev. Neuropsychol.
Perception of complex sounds: abnormal pattern of cortical activation in autism
Am. J. Psychiatry
MP RAGE: a three-dimensional, t1-weighted, gradient-echo sequence–initial experience in the brain
Radiology
Pervasive rightward asymmetry shifts of functional networks in autism spectrum disorder
JAMA Psychiatry
Prevalence of autism spectrum disorder among children aged 8 years-autism and developmental disabilities monitoring network, 11 sites, United States, 2010
Morbidity and Mortality Weekly Report
LIBSVM: a library for support vector machines
ACM Trans. Intell. Syst. Technol.
DPARSF: a MATLAB toolbox for “pipeline” data analysis of resting-state fMRI
Front. Syst. Neurosci.
Diagnostic classification of intrinsic functional connectivity highlights somatosensory, default mode, and visual regions in autism
NeuroImage
Functional connectivity in a baseline resting-state network in autism
Neuroreport
A whole brain fmri atlas generated via spatially constrained spectral clustering
Hum. Brain Mapp.
Aberrant striatal functional connectivity in children with autism
Biol. Psychiatry
The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism
Mol. Psychiatry
Cited by (76)
Automatic characterization of cerebral MRI images for the detection of autism spectrum disorders
2024, Intelligence-Based MedicineSpectral representation of EEG data using learned graphs with application to motor imagery decoding
2024, Biomedical Signal Processing and ControlFacial Image-Based Autism Detection: A Comparative Study of Deep Neural Network Classifiers
2024, Computers, Materials and Continua