Skip to Content
You are currently on the new version of our website. Access the old version .
DiagnosticsDiagnostics
  • Article
  • Open Access

16 March 2023

20 Pages

A Model for Diagnosing Autism Patients Using Spatial and Statistical Measures Using rs-fMRI and sMRI by Adopting Graphical Neural Networks

and
Department of Computer Science and Engineering, Sri Sivasubramaniya Nadar College of Engineering, Chennai 603110, India
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.

Abstract

This article proposes a model to diagnose autism patients using graphical neural networks. A graphical neural network relates the subjects (nodes) using the features (edges). In our model, radiomic features obtained from sMRI are used as edges, and spatial-temporal data obtained through rs-fMRI are used as nodes. The similarity between first-order and texture features from the sMRI data of subjects are derived using radiomics to construct the edges of a graph. The features from brain summaries are assembled and learned using 3DCNN to represent the features of each node of the graph. Using the structural similarities of the brain rather than phenotypic data or graph kernel functions provides better accuracy. The proposed model was applied to a standard dataset, ABIDE, and it was shown that the classification results improved with the use of both spatial (sMRI) and statistical measures (brain summaries of rs-fMRI) instead of using only medical images.

1. Introduction

Autism spectrum disorder (ASD) is a neuro-developmental disability that affects the social and behavioural abilities of an individual. People with ASD find it challenging to interact with society and accept changes in their routine work [1,2]. The diagnosis of autism includes observations made by (1) paediatricians, (2) child psychologists, (3) neurologists, and (4) therapists. To analyse the behavioural pattern of a child, visits to a psychologist and paediatrician is needed, which can take from months to years. Some distinguishable findings of autism-affected brains are (i) an increased brain volume, (ii) reduced functional connectivity in the temporal lobe and visual cortex, and (iii) more folds in the left parietal, temporal and frontal lobes, areas that affect social behaviour and language production. Magnetic resonance imaging (MRI), structural MRI (sMRI), functional MRI (fMRI) and resting state fMRI (rs-fMRI) help with the early and accurate diagnosis of autism. sMRI data provide structural information about the brain, such as its size, volume, convolutions, grey matter density and white matter distribution. fMRI and rs-fMRI provide information about functional dependencies in brain regions, which can aid in the identification of a child’s social and behavioural activities [3,4]. The brain’s major characteristics can be determined by means of sMRI and rs-fMRI.
Radiomics is used in medical imaging to augment the available data by applying a mathematical analysis to quantitative features and extracting pixel correlations and the distribution of signal intensities in brain regions (voxels). The features obtained through radiomics are (i) shape and size-based, (ii) image density histograms, (iii) correlations between voxels, (iv) texture from filtered images, and (v) form factor features [5,6,7].
Given a graph G ( V , E ) as the input, the graph convolution network (GCN) takes an input as a matrix X ( N × F ) , where N is the number of nodes and F is the features of a node, and an adjacency matrix A ( N × N ) . A degree matrix D is a diagonal matrix that stores the sum of rows of an adjacency matrix A in the diagonal D i = ∑ j · A i j . Each node in the graph has a d dimensional feature vector, so the entire feature matrix can be described as X ∈ R n × d . All nodes in the graph belong to one of the c classes specified in the c dimensional vector y i ∈ 0 ,   1 ,   … , c .
A GCN [8,9] is similar to a neural network learning new features of each node over multiple layers ( refer Figure 1). The initial node representation is H 0 = X , the input to the GCN. In a k-layer GCN, the hidden layer representation is averaged with its neighbours at the beginning of each layer. A three-step process of nonlinear propagation, linear transformation and nonlinear activation occurs in each hidden layer.
Figure 1. GCN Architecture.
  • Feature propagation involves averaging the feature of each node v i at the beginning of layer H i with its local neighbourhood, as shown in Equation (1).
    H i = ∑ j = 1 n A i j ( d i + 1 ) ( d j + 1 )
    The normalised matrix representation is specified in Equation (2)
    H i = 1 D ¯ · A ¯ 1 D ¯ H i − 1 · W i − 1
    where A ¯ = A + I N and D ¯ = ∑ j A i j ¯ is the Degree matrix of A.
  • Feature and nonlinear transformations associate each layer with a learned weight matrix W i , and hidden layer transformations are transformed linearly. A feature representation of layer H k is generated after applying ReLU. This step is followed by feature propagation for the next layer. Equation (3) provides the criteria for updating the k th layer [10].
    H k ← R e L U H k , W k
  • The classification of a node in a k th layer GCN is obtained using Equation (4).
    Y G C N = S o f t m a x ( H k )
Analyses using brain summary methods typically involve the extraction of statistical features from the rs-fMRI data. These features can be used to quantify the strength and direction of functional connectivity between different brain regions to characterize the magnitude and consistency of the rs-fMRI signal within a given brain region. These statistical features can provide information on the underlying physiological processes and can be used to identify abnormalities in brain function. Brain summary measures are used to reduce the 4 D time series data using statistical measures into a single number per voxel. The summary measures chosen are listed below:
  • First, the Regional Homogeneity (ReHo) uses Kendall’s coefficient of concordance to calculate the similarity between the time series of a voxel and its neighbour.
  • Second, the Amplitude of Low-Frequency Fluctuations (ALFF) uses the amplitude of each region of interest (ROI) in the frequency band of 0.1 to 0.08 Hz. It shows the spontaneous neuro behaviour in rs-fMRI data. The power spectrum is obtained from the time series signal by a fast Fourier transform. The averaged square root value of each voxel in the power spectrum across the frequency range is considered the ALFF value.
  • Third, the Fractional Amplitude of Low-Frequency Fluctuations (fALFF) finds the power ratio in the frequency band relative to the full frequency band of 0 to 0.25 Hz.
  • Fourth, the Degree Centrality ( D C b i n , D C w e i g h t e d ) measures the edge count of a region. It specifies the number of regions to which the current node is related. This measure proves higher associations in cortical regions.
  • Fifth, the Eigenvector Centrality ( E C b i n , E C w e i g h t e d ) provides the centrality of a node based on the number of edges connecting central nodes; i.e., a node is important if it is connected to more important nodes.
  • Sixth, the Local Functional Connectivity Density (LFCD) calculates the local functional connection of a voxel within a ROI and the functional connections of voxels between hemispheres.
  • Seventh, the Voxel-Mirrored Homotopic Connectivity (VMHC) calculates the voxel-wise connectivity between symmetrical brain hemispheres. ROIs that show high connectivity between the left hemisphere, and its right mirrored counterpart have high VHMC values, as shown in Figure 2.
  • Eighth, dual regression uses the independent component group maps obtained from the independent component analysis (ICA) as network templates. A spatial regressor uses the spatial map to identify the time series associated with voxels on the corresponding map, followed by a temporal regressor, which uses the time series to fetch the complete set of voxels activated in that time series. The result of these steps is a subject-specific spatial map based on the original spatial map.
Figure 2. Comparison of Mirrored Voxels across Hemispheres in a Symmetrical Brain using VMHC.
The motivation behind the proposed work was to use both structural features and functional connectivity to contribute to a study of the brain. The summary measures provide the most essential statistical features of the brain. More accurate results could be obtained by fusing more summaries instead of obtaining images from a single summary measure.
The highlights of our proposed model are as follows:
  • Radiomic features obtained from sMRI using the Destrieux atlas were used to identify similarities between individuals (nodes) to construct an edge for the graph. These features can capture information about various aspects of the image, such as the texture, shape, size, intensity, and spatial relationship.
  • We extracted 48,544 radiomic features passed to stacked autoencoders for a dimensionality reduction. We obtained a similarity index between the 150 features of each node using an improved sqrt cosine similarity function to draw an edge between the nodes.
  • Brain summaries provide spatial features passed to the Multichannel CNN to obtain the feature vector for each node. We obtained 3D volume images for each brain summary derivative and combined them on the fourth (channel) axis. The multichannel 3D CNN produces a feature vector with a size of 1024.
  • Population Graphs were constructed with each subject as a node and edges based on the similarities identified in the brain’s structure. The consideration of structural similarities reduced the heterogeneity of the ABIDE dataset, as MRI data were collected from 16 locations.
  • Graph convolution networks were used to produce complete graphs, which helped to classify nodes as ASD or Typical Controls (TC).
The rest of the paper is organized as follows: Section 2 discusses the work related to the GCN and deep learning methods in terms of classifying ASD from TC. Section 3 elaborates on the proposed method, Section 4 explores the experimental results, Section 5 details the research discussion, and finally, Section 6 concludes the paper with possible future work.

3. Materials and Methods

In this section, we present an overview of the proposed model, followed by an explanation of the proposed model for the classification of a subject as having ASD or being a healthy control given sMRI and rs-fMRI data based on our graph convolution model.

3.1. Proposed Method

The workflow of the proposed model using a graph convolutional network for classifying ASD from TC is illustrated in Figure 3, which uses both the structural and functional connectivity of the brain. The morphological abnormalities observed in the brains of ASD patients have variations in grey and white matter volumes and the sizes of cerebral, cerebellar, and subcortical regions. The intrinsic functional connectivity between the regions of the brain was identified using rs-fMRI.
Figure 3. Workflow of Proposed Model. (1: ROI Extraction, 2: Features through Fischer’s Score, 3: Features through Stacked Autoencoder, 4: Builing Adjacency Matrix, 5: Extraction of Summary Measures, 6: Building Feature Vector for Each Node, 7: Constructing Population Graph, 8: Complete Graph using GCN).
The sMRI image obtained from the dataset is passed to the Deistreiux atlas for ROI extraction (step 1). A set of quantitative features is extracted from the segmented regions of the sMRI image. These features are classified into different groups, such as intensity-based features, texture-based features, and shape-based features. The features collected through radiomics are ranked using Fischer’s score, and the top 2000 features are selected (step 2). They are passed to stacked autoencoders to extract the top 150 features for each subject (step 3). The Sqrt-cosine similarity [25] is calculated between the nodes to identify whether an edge can be drawn between them (step 4).
Brain summary methods are commonly used in rs-fMRI analyses to summarize the functional connectivity patterns within the brain [26]. These methods involve the extraction of summary measures from rs-fMRI that capture the connectivity between different brain regions, thus reducing the temporal component of the 4D rs-fMRI image (step 5). Each brain summary method produces a 3D volume of a rs-fMRI image, consisting of 19 volumes of 3D images. They are passed to the multichannel 3D CNN, which yields 1024 features for each subject (step 6). Then, the complete dataset can be represented as a graph, where each subject is a node with 1024 features, and the similarity measure can be used to determine the edges between other nodes (step 7). They are then passed to the GCN, which produces a complete graph that helps to classify an unidentified node (step 8).

3.2. Building the Adjacency Matrix

The edges between the nodes of the graph are identified using the adjacency matrix. Then, the radiomic features obtained from sMRI are used. Radiomic features are used to better understand the textural and volumetric features of brain regions. Py-Radiomics (refer Figure 4) is used to extract the radiomic features from the preprocessed sMRI images. The wavelet transform in Py-Radiomics uses the Daubechies wavelet filter as the default wavelet filter. The high-pass filter of the Daubechies wavelet filter is applied in 8 different directions, corresponding to the 8 taps or coefficients of the filter. Therefore, the wavelet transform at level 1 in Py-Radiomics produces a set of sub-bands, each corresponding to a different frequency band and spatial orientation. After the wavelet transform has been performed, the resulting wavelet coefficients are used as the input to a feature extractor. This calculates a large number of features from the wavelet coefficients, including first-order statistics and texture features.
Figure 4. Radiomics Feature Extraction.
The Destrieux atlas, also known as the Cortical Parcellation atlas, is used to capture the structural information, such as the cortical thickness, grey matter volume, and curvature, calculated from the MRI images. The Destrieux atlas selects 74 regions from the right hemisphere and 74 regions from the left hemisphere, giving a total of 148 regions of interest (ROI). We obtain 19 histogram-based features, 22 texture-based (GLCM) features, and wavelet-based features in 8 directions. Overall, we have 328 features i.e., ( 19 + 22 ) × 8 .
The wavelet features are obtained from all possible directions of Low (L) and High (H) pass filters. Table A1 provides the radiomic features extracted with the help of Py-radiomics [27], and Table A2 specifies the terms used. We obtain 328 features for each ROI; for each subject, there will be 48,544 features, i.e., ( 328 × 148 ) . Fisher’s score is a supervised algorithm that is used to select features (refer to Algorithm 1 and Table A3), and it returns each variable’s rank based on its score in descending order.
Algorithm 1: Feature Ranking using Fischer’s Score
Diagnostics 13 01143 i001
    Two thousand features are selected using Fisher’s score and passed to a stacked autoencoder (4 levels) for dimensionality reduction. Stacked autoencoders produce a latent space that is less sparse, more predictable, continuous [23,24] and yields 150 features. Algorithm 2 specifies the encoding process of the stacked autoencoder.
Algorithm 2: Stacked Autoencoder
Diagnostics 13 01143 i002
The level of similarity between nodes is identified using the improved sqrt-cosine measurement, as mentioned in Equation (5), and is based on the Hellinger distance. The main advantage of the Hellinger distance over the Euclidean or Manhattan distance sis that it is better suited for comparing probability distributions. The Euclidean and Manhattan distances are commonly used in traditional distance metrics and are useful for measuring distances between the vectors of numerical values.
S i m ( x , y ) = ∑ i = 1 150 x i · y i ∑ i = 1 150 x i · ∑ i = 1 150 y i
However, they are not well suited for measuring distances between probability distributions, because they do not take into account the structure of probability distributions, which is fundamentally different from that of vectors. The Hellinger distance, on the other hand, is specifically designed for comparing probability distributions and takes into account the structure of the distribution. It is defined as the square root of the sum of the squared differences between the square roots of the probabilities. This metric is useful for comparing distributions because it is bounded between 0 and 1, with 0 representing identical distributions and 1 representing completely different distributions. As it is similarity-based, it is associated with values between 0 and 1, which could be better associated with probability-based approaches. Thus, every node is connected to at least one other node, maximizing the sparsity. The threshold is set as the minimum of all maximum values in a row.

3.3. Feature Selection

rs-fMRI is used to identify the brain connectivity by calculating the blood oxygen levels, where the higher the oxygen level in a brain region is, the greater the activation of the region is [28]. We use brain summaries that are statistical derivatives to eliminate the temporal component of the MRI scan [29] and 3D CNN to extract features from them (refer Figure 5).
Figure 5. Feature Selection from rs-fMRI images.
For each subject, the ten brain summary derivatives are calculated from the 4D fMRI data, resulting in a 3D volume brain image of 61 × 73 × 61 , meaning that we have a stack of 61 slices of 61 × 73 2D images. The spatial characteristics of the brain image, such as the brain regions, volume, cortical folding, and white matter tracts, are restored. Each volume can be viewed as axial, sagittal, or coronal.
For the dual regression derivative, ten different volumes, 61 × 73 × 61 in size, (10 spatial group maps from the Independent Component Analysis (ICA)) enhance the contrast of the brain tissue and separate unknown signal sources into statistically independent components. Maps 1, 2 and 3 strongly correspond to the visual behavioural domain; Map 4 is the default mode network; Map 5 corresponds to the cerebellar region; Map 6 contains sensorimotor information; Map 7 includes auditory paradigms, Map 8 corresponds to cognition and Maps 9 and 10 are left–right mirrors, which correspond to the perception and language paradigm). In total, there are 19 3D brain volumes for each subject. A fourth dimension, the channel axis, concatenates different derivatives as a multichannel. Dual regression produces images, 61 × 61 × 73 × 10 in size (10 denotes the channel for each map). When the ALFF needs to be concatenated with a dual regression to form a new multichannel, we add a channel axis, so its dimensions become 61 × 61 × 73 × 1 . Now the concatenation is performed on the channel axis, resulting in new multichannel data with dimensions of 61 × 61 × 73 × 11 .
We build a 3D convolution model to obtain features from the multichannel brain volumes. A convolution block is characterized as a convolution layer with ReLU activation followed by batch normalization and max-pooling. The Kernel size is set as ( 3 × 3 × 3 ) , and the pool size is ( 2 × 2 × 2 ) . The second convolution block is repeated four times, starting with 32 filters and doubling at each block, so the last convolution block has 256 filters. The global average pooling layer generates channel descriptors by 3D pooling of all previous layers. The flattened vector is given as the input to a dense layer consisting of 1024 nodes. Thus, 1024 features are obtained for each node by using the multichannel 3D CNN.

3.4. Graph Convolution

The population graph is constructed as nodes with 1024 features each, and the edges are determined with their similarity measured by the radiomic features of the sMRI data. In the GCN-based model, as specified in Figure 6, features are used to characterize the nodes, while edges denote the associations between the nodes [30,31]. The graph convolutional network uses two hidden layers, with each node containing 16 units with dropouts of 0.005 and 0.4, respectively.
Figure 6. Graph Convolution.
In contrast to similar work carried out previously, in nongraphical-based methods, the correlation coefficients obtained from the functional connectivity matrix are used. The ASD-Diagnet [15,16] uses functional connectivity, a measure of the temporal correlations between different brain regions. The method has been used extensively to study the neural basis of ASD. Correlation coefficients can be influenced by head motion, the signal-to-noise ratio, and individual differences in brain anatomy. Failure to account for these confounding factors may result in errors or a reduced classification accuracy. Multimodal data classification [17] is performed by ensembling the classification results of different modalities by either blending or voting. Different modalities are only ensembled after classification. Ensembling multiple modalities before classification can improve the accuracy compared to ensembling the classification results. The accuracy is improved because each modality can capture different aspects of the data, and by combining them, a more comprehensive representation of the data can be created.
In graph-based methods, s-GCN [22] creates a graph with phenotypic data (edges) and functional connectivity features from rs-fMRI data used as nodes. The use of phenotype data to construct edges might lead to a lack of heterogeneity, since phenotype data do not consider variations in topographical locations. The MVS-GCN [19] generates multiple view graphs from brain subnetworks, which may result in redundant feature learning. The heterogeneity of the brain is not considered when forming subnetworks, which may result in a lack of functional connectivity between regions across subnetworks. Eigen-based decomposition pooling, as used in EigenGCN [18], causes a loss of crucial information about features due to the presence of multiple pooling layers. In our model, we considered both the structural and functional connectivity of brain data and used both to construct a graph for a subject, enhancing the classification accuracy. Table 2 highlights the findings of similar recent models and compares them with the proposed model.
Table 2. Comparison of Proposed Model with Recent Similar Models.

4. Experimental Results

We used the Ubuntu operating system, an Intel i9 12900 processor, 16 GB RAM, and an Nvidia RTX A2000 GPU for all experiments conducted.

4.1. Dataset

The ABIDE is a preprocessed dataset used for autism brain imaging research. The dataset includes data from 16 sites, consisting of sMRI, rs-fMRI, and phenotypic information from 539 autistic patients and 573 typical controls. The data were harmonized and standardized to ensure consistent representation across all sites. The specific preprocessing steps performed on the dataset included the following:
  • Motion correction: To reduce the impact of head movement on the fMRI data, the images were corrected for motion using tools such as MCFLIRT and MotionCorr.
  • Spatial normalisation: To ensure a consistent representation of brain structures across participants, the data were transformed into a shared space using a standard brain template ( Montreal Neurological Institute template).
  • Noise reduction: Various noise sources, such as physiological noise from the heart and respiratory system, were removed from the data using physiological regression or CompCor.
  • Data cleaning: The data were checked for quality and outliers, and any poorly performing time points or participants were removed from the dataset.
Functional and structural preprocessing and the calculation of cortical measures were carried out using well-defined pipelines.

4.2. Parameter Search

The hyperparameters used for CNN models include the number of convolutional layers, the number of filters, the dropout rate, and the number of dense nodes in a dense layer. Various values that were used are mentioned in Table 3.
Table 3. Hyperparameter Selection.
With the above hyperparameter values for CNN, a better accuracy score of 69.45 (Figure 7) was obtained with a 3D CNN model (number of convolutional blocks after the first layers = 4, filters = 64, dropout rate = 0.5, number of nodes in a dense layer = 1024).
Figure 7. Parameter Selection for the 3D CNN.

4.3. Experiments

The population graph of a graph convolution model is formed with the similarity of features extracted from sMRI radiomics, and its edges contain the characteristics (1024) of the 19 retrieved derivatives. To demonstrate the significance of each summary measure and the useful outcome produced by combining the summary measurements, we conducted our experiment individually for each brain summary. Despite the variety in the topographical locations of the data collected, our model produced better results. The majority of deep learning models that use fMRI to diagnose autism have higher local accuracy levels. Our GCN model offers improved accuracy for multisite data in ABIDE, because it creates an edge between nodes based on similarities in the sMRI data.

4.4. Evaluation

The accuracy, specificity and sensitivity are commonly used metrics to evaluate the performance of deep learning models and can help to determine their efficiency.
  • The accuracy measures the proportion of correct predictions made by the model. It is a useful metric to evaluate the model’s overall performance, but it may not be the best metric for imbalanced datasets.
    A c c u r a c y = ( T P + T N ) ( T P + T N + F P + F N ) · 100
  • The sensitivity (also called the recall or true positive rate) measures the proportion of actual positive cases correctly identified by the model. It helps to detect false negatives and is essential in applications where the identification of positive cases is critical.
    S e n s i t i v i t y = ( f P + F N ) ( T P + T N + F P + F N ) · 100
  • The specificity (also called the true negative rate) measures the proportion of actual negative cases correctly identified by the model. It helps to detect false positives and is essential in applications where avoiding false alarms is critical.
    S p e c i f i c i t y = T P ( T P + F P ) · 100
  • The precision measures the accuracy of positive predictions made by the model. It tells us how many of the positive predictions made by the model were actually correct.
    P r e c i s i o n = T N ( F P + T N ) · 100
By examining the values of these metrics, we can gain insight into the strengths and weaknesses of the model’s performance (refer Figure 8). The standard deviations of these metrics for the proposed model combining 19 derivatives were 0.42 % , 0.44 % , 0.13 % , and 0.48 % , respectively.
Figure 8. Accuracy of the Proposed Model.

5. Discussion

Most deep learning models use fMRI images to classify ASD patients and controls. The integration of structural and functional features provides a whole-brain analysis for better feature identification. Constructing a graph model based on the relationships among the sMRI features obtained using radiomics provides better relationships between individuals than using phenotypic data, as used in previous works. We performed experiments by using and ensembling the summary measures. Combining features extracted from various brain summaries provides a better classification result than applying a single summary measure. We integrated data from all sites and topographical locations in the ABIDE repository to ensure that the results were independent of demographic changes. Our model produced a better accuracy instead of leaving out a site strategy. The use of a graph convolutional model was shown to provide better results by grouping nodes based on their structural similarities.

Comparison with the State-of-the-Art Methods

To demonstrate the overall performance of our brain network in the diagnosis of ASD, we compared our model with various state-of-the-art techniques. In addition, we compared our model with nongraph and GCN-based models to provide a comprehensive study.
To extract features and improve the classification performance, ASD-DiagNet uses a single-layer perceptron (SLP) and an autoencoder that are jointly learned. In contrast, when using the Restricted Boltzmann Machine, DBN features are learned from rs-fMRI and sMRI data (grey matter and white matter).
Based on the phenotypic data from the ABIDE dataset, the Siamese GCN (sGCN) learns patterns related to the similarities between graph nodes. In contrast, the EigenGCN operates on the eigenvectors of the adjacency matrix, which are used to encode the graph’s structure and related information. This makes it possible to run graph convolutions directly, rather than using the original graph representation.
A comparison of the efficiency of our model compared with various state-of-the-art models is presented in Table 4. ASD DiagNet uses Pearson’s correlation measure to determine the functional connectivity from the functional MRI data. Nineteen thousand, nine hundred features were selected, which was insufficient for training a deep learning model, so the Synthetic Minority Over-sampling Technique (SMOTE) was used for augmentation purposes. In our method, we merged the sMRI radiomic features with features extracted from the 3D CNN (48,544 + 1024 = 49,568) as the input for the autoencoders. The performance of the model was improved by 8.34 % . Similarly, the merged features were fed to the DBN model, which uses a hierarchically restricted Boltzmann machine for feature learning. In this method, the average signals of each region of interest are calculated using the AAL atlas (116 ROI) and the grey and white matter volumes of the sMRI data. Our feature vector outperformed the model by producing an accuracy of 82.45 % with a depth of 3. Considering graph-based networks, our model was proven to produce better results than the s-GCN and EigenGCN. The use of structural similarities to construct the edges of a graph was shown to be more efficient than using phenotypic data or graph kernels.
Table 4. Comparison of the Proposed Model with State-of Art Models.

6. Conclusions

We designed and implemented a graph convolutional model to classify ASD using structural and rs-fMRI data information. The graphical model provides a better association between subjects, which helps us to classify them better. It is evident that fusing structural and rs-fMRI data helps to generate a well-learned graph and improves the classification accuracy. Our model is based on the entire brain’s functional connectivity and the brain’s grey matter, structure and shape. This distinguishes the behavioural, linguistic and social characteristics of patients with ASD from healthy controls. We determined the similarity between patients based on the structures of their brains, which proved to be a better similarity measure than phenotypical data, such as age, gender and location. Every model has some limitations. In our model, these are as follows: (i) the number of steps processed is higher when compared to those of other GCN models (EigenGCN and S-GCN); (ii) it takes more time to form the graph due to the formation of edges from the sMRI data; and (iii) the model is more complex due to the use of both sMRI and rs-fMRI data. Moreover, we used only the ABIDE dataset, and with a little pre-processing, our model could be tested on other MRI datasets such as the NDAR, open fMRI, ADNI, etc. Future research should consider the fusion of the INCLEN diagnostic dataset and videos captured during psychologists’ visits and MRI images of ASD patients to produce a better classification accuracy.

Author Contributions

Conceptualization, K.M. and S.J.; Methodology, K.M.; Software, K.M.; Formal analysis, K.M.; Writing—original draft, K.M.; Writing—review & editing, S.J.; Visualization, K.M. and S.J.; Supervision, S.J. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not Applicable.

Data Availability Statement

The data used in this study is from public database ABIDE. The data is available at http://fcon_1000.projects.nitrc.org/indi/abide/.

Conflicts of Interest

The authors declare no conflict of interest.

Abbreviations

The following abbreviations are used in this manuscript:
ASDAutism Spectrum Disorder
TCTypical Controls
ABIDEAutism Brain Imaging Data Exchange
MRIMagnetic Resonance Imaging
sMRIStructural MRI
fMRIFunctional MRI
rs-fMRIResting State Functional MRI
CNNConvolutional Neural Network
GCNGraph Convolutional Network
ROIRegion of Interest
ReHoRegional Homogeneity
ALFFAmplitude of Low Frequency Fluctuations
fALFFFunctional Amplitude of Low Frequency Fluctuations
DCDegree Centrality
ECEigen vector Centrality
LFCDLocal Functional Connectivity Density
VMHCVoxel-Mirrored Homotopic Connectivity
DRDual Regression
ICAIndependent Component Analysis
BOLDBlood Oxygenation Level Dependent
GLCMGray Level Co-occurrence Matrix

Appendix A

Table A1. Features Extracted from PY-Radiomics.
Table A2. Terms used in Radiomic Feature extraction Algorithms.
Table A3. Terms used in Fischer Score Claculation.

References

  1. Cheng, W.; Rolls, E.T.; Zhang, J.; Sheng, W.; Ma, L.; Wan, L.; Luo, Q.; Feng, J. Functional connectivity decreases in autism in emotion, self, and face circuits identified by Knowledge-based Enrichment Analysis. Neuroimage 2017, 148, 169–178. [Google Scholar] [CrossRef] [Scilit]
  2. Hazlett, H.C.; Gu, H.; Munsell, B.C.; Kim, S.H.; Styner, M.; Wolff, J.J.; Elison, J.T.; Swanson, M.R.; Zhu, H.; Botteron, K.N.; et al. Early brain development in infants at high risk for autism spectrum disorder. Nature 2017, 542, 348–351. [Google Scholar] [CrossRef] [Scilit]
  3. Giangiacomo, E.; Visaggi, M.C.; Aceti, F.; Giacchetti, N.; Martucci, M.; Giovannone, F.; Valente, D.; Galeoto, G.; Tofani, M.; Sogos, C. Early Neuro-Psychomotor Therapy Intervention for Theory of Mind and Emotion Recognition in Neurodevelop-mental Disorders: A Pilot Study. Children 2022, 9, 1142. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  4. Kong, Y.; Gao, J.; Xu, Y.; Pan, Y.; Wang, J.; Liu, J. Classification of autism spectrum disorder by combining brain connectivity and deep neural network classifier. Neurocomputing 2018, 324, 63–68. [Google Scholar] [CrossRef] [Scilit]
  5. Cao, X.; Wang, X.; Xue, C.; Zhang, S.; Huang, Q.; Liu, W. A Radiomics Approach to Predicting Parkinson’s Disease by Incor-porating Whole-Brain Functional Activity and Gray Matter Structure. Front. Neurosci. 2020, 14, 751. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  6. Donisi, L.; Cesarelli, G.; Castaldo, A.; De Lucia, D.R.; Nessuno, F.; Spadarella, G.; Ricciardi, C. A Combined Radiomics and Machine Learning Approach to Distinguish Clinically Significant Prostate Lesions on a Publicly Available MRI Dataset. J. Imaging 2021, 7, 215. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  7. Dekhil, O.; Hajjdiab, H.; Shalaby, A.; Ali, M.T.; Ayinde, B.; Switala, A.; Elshamekh, A.; Ghazal, M.; Keynton, R.; Barnes, G.; et al. Using resting state functional MRI to build a personalized autism diagnosis system. PLoS ONE 2018, 13, e0206351. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  8. Kumar, A.; Tripathi, A.R.; Satapathy, S.C.; Zhang, Y.-D. SARS-Net: COVID-19 detection from chest x-rays by combining graph convolutional network and convolutional neural network. Pattern Recognit. 2021, 122, 108255. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  9. Ghorbani, M.; Kazi, A.; Baghshah, M.S.; Rabiee, H.R.; Navab, N. RA-GCN: Graph convolutional network for disease predic-tion problems with imbalanced data. Med. Image Anal. 2021, 75, 102272. [Google Scholar] [CrossRef] [Scilit]
  10. Yu, Z.; Huang, F.; Zhao, X.; Xiao, W.; Zhang, W. Predicting drug–disease associations through layer attention graph convolu-tional network. Brief. Bioinform. 2020, 22, bbaa243. [Google Scholar] [CrossRef] [Scilit]
  11. Ahammed, S.; Niu, S.; Ahmed, R.; Dong, J.; Gao, X.; Chen, Y. DarkASDNet: Classification of ASD on Functional MRI Using Deep Neural Network. Front. Neuroinform. 2021, 15, 635657. [Google Scholar] [CrossRef] [Scilit]
  12. Haweel, R.; Seada, N.; Ghoniemy, S.; Alghamdi, N.S.; El-Baz, A. A CNN Deep Local and Global ASD Classification Ap-proach with Continuous Wavelet Transform Using Task-Based FMRI. Sensors 2021, 21, 5822. [Google Scholar] [CrossRef] [Scilit]
  13. Li, X.; Dvornek, N.C.; Papademetris, X.; Zhuang, J.; Staib, L.H.; Ventola, P.; Duncan, J.S. 2-Channel convolutional 3D deep neural network (2CC3D) for fMRI Analysis: ASD classification and feature learning. In Proceedings of the IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), Washington, DC, USA, 4–7 April 2018; pp. 1252–1255. [Google Scholar] [CrossRef] [Scilit]
  14. Leming, M.J.; Baron-Cohen, S.; Suckling, J. Single-participant structural similarity matrices lead to greater accuracy in classi-fication of participants than function in autism in MRI. Mol. Autism. 2021, 12, 1–15. [Google Scholar] [CrossRef] [Scilit]
  15. Eslami, T.; Mirjalili, V.; Fong, A.; Laird, A.R.; Saeed, F. ASD-DiagNet: A Hybrid Learning Approach for Detection of Autism Spectrum Disorder Using fMRI Data. Front. Neuroinformatics 2019, 13, 70. [Google Scholar] [CrossRef] [Scilit]
  16. Sewani, H.; Kashef, R. An Autoencoder-Based Deep Learning Classifier for Efficient Diagnosis of Autism. Children 2020, 7, 182. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  17. Ali, L.; He, Z.; Cao, W.; Rauf, H.T.; Imrana, Y.; Bin Heyat, B. MMDD-Ensemble: A Multimodal Data–Driven Ensemble Ap-proach for Parkinson’s Disease Detection. Front. Neurosci. 2021, 15, 754058. [Google Scholar] [CrossRef] [Scilit]
  18. Ma, Y.; Wang, S.; Aggarwal, C.C.; Tang, J. Graph Convolutional Networks with Eigen Pooling. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery Data Mining (KDD 19), Anchorage, AK, USA, 4–8 August 2019; Association for Computing Machinery: New York, NY, USA, 2019; pp. 723–731. [Google Scholar] [CrossRef] [Scilit]
  19. Wen, G.; Cao, P.; Bao, H.; Yang, W.; Zheng, T.; Zaiane, O. MVS-GCN: A prior brain structure learning-guided multi-view graph convolution network for autism spectrum disorder diagnosis. Comput. Biol. Med. 2022, 142, 105239. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  20. Song, X.; Zhou, F.; Frangi, A.F.; Cao, J.; Xiao, X.; Lei, Y.; Wang, T.; Lei, B. Graph convolution network with similarity awareness and adaptive calibration for disease-induced deterioration prediction. Med. Image Anal. 2020, 69, 101947. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  21. Lei, B.; Zhu, Y.; Yu, S.; Hu, H.; Xu, Y.; Yue, G.; Wang, T.; Zhao, C.; Chen, S.; Yang, P.; et al. Multi-scale enhanced graph convolutional network for mild cognitive impairment detection. Pattern Recognit. 2023, 134, 109106. [Google Scholar] [CrossRef] [Scilit]
  22. Parisot, S.; Ktena, S.I.; Ferrante, E.; Lee, M.; Guerrero, R.; Glocker, B.; Rueckert, D. Disease prediction using graph convolutional networks: Application to Autism Spectrum Disorder and Alzheimer’s disease. Med. Image Anal. 2018, 48, 117–130. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  23. Shi, D.; Zhang, H.; Wang, G.; Wang, S.; Yao, X.; Li, Y.; Guo, Q.; Zheng, S.; Ren, K. Machine Learning for Detecting Parkinson’s Disease by Resting-State Functional Magnetic Resonance Imaging: A Multicenter Radiomics Analysis. Front. Aging Neurosci. 2022, 14, 806828. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  24. Wang, L.; Feng, Q.; Ge, X.; Chen, F.; Yu, B.; Chen, B.; Liao, Z.; Lin, B.; Lv, Y.; Ding, Z. Textural features reflecting local activi-ty of the hippocampus improve the diagnosis of Alzheimer’s disease and amnestic mild cognitive impairment: A radiomics study based on functional magnetic resonance imaging. Front. Neurosci. 2022, 16, 970245. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  25. Sohangir, S.; Wang, D. Improved sqrt-cosine similarity measurement. J. Big Data 2017, 4, 25. [Google Scholar] [CrossRef] [Scilit]
  26. Dekhil, O.; Ali, M.; El-Nakieb, Y.; Shalaby, A.; Soliman, A.; Switala, A.; Mahmoud, A.; Ghazal, M.; Hajjdiab, H.; Casanova, M.F.; et al. A Personalized Autism Diagnosis CAD System Using a Fusion of Structural MRI and Resting-State Functional MRI Data. Front. Psychiatry 2019, 10, 392. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  27. Py-Radiomic Features. Radiomic Features. Available online: https://pyradiomics.readthedocs.io/en/latest/features.html (accessed on 20 October 2022).
  28. Howsmon, D.P.; Kruger, U.; Melnyk, S.; James, S.J.; Hahn, J. Classification and adaptive behavior prediction of children with autism spectrum disorder based upon multivariate data analysis of markers of oxidative stress and DNA methylation. PLoS Comput. Biol. 2017, 13, e1005385. [Google Scholar] [CrossRef] [Scilit]
  29. Williams, B.; Kabbage, M.; Kim, H.J.; Britt, R.; Dickman, M.B. Tipping the balance: Sclerotinia sclerotiorum secreted oxalic acid suppresses host defenses by manipulating the host redox environment. PLoS Pathog. 2011, 7, e1002107. [Google Scholar] [CrossRef] [Scilit]
  30. Nie, W.; Chang, R.; Ren, M.; Su, Y.; Liu, A. I-GCN: Incremental Graph Convolution Network for Conversation Emotion Detection. IEEE Trans. Multimedia 2021, 24, 4471–4481. [Google Scholar] [CrossRef] [Scilit]
  31. Chu, Y.; Wang, G.; Cao, L.; Qiao, L.; Liu, M. Multi-Scale Graph Representation Learning for Autism Identification with Func-tional MRI. Front. Neuroinform. 2022, 15, 802305. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.