Automatic diagnosis of autism spectrum disorders in children through resting-state functional magnetic resonance imaging with machine vision

Introduction

Autism spectrum disorder (ASD) is a group of developmental disorders affecting the nervous system. Its main characteristics include impairments in social interactions, communication, repetitive behaviors and limited interests (1-3). Numerous studies have investigated the components of the neural systems linked to clinical symptoms in ASD. Certain key areas of the brain, including the frontotemporal lobe, frontoparietal cortex, amygdala, hippocampus, basal ganglia, and anterior cingulate cortex (ACC), have been proposed to play a role in the manifestation of ASD symptoms (4). Abnormalities in the inferior frontal gyrus (IFG), superior temporal sulcus (STS), and Wernicke’s area could affect social language processing and attention. The frontal lobe, superior temporal cortex, parietal cortex, and amygdala are involved in social behavior impairments. The orbitofrontal cortex (OFC) and caudate nucleus are linked to repetitive behaviors in individuals with ASD (5).

Clinical evaluations for ASD involve assessments like ADOS and ADI-R scores, IQ tests, Social Response Scale (SRS) and the Vinland Behavioral Adjustment Scale (VABS) (6). However, these assessments are time consuming and require expertise. Imaging techniques can serve as valuable diagnostic tools for detection ASD and can supplement the clinical diagnosis. Among these techniques, magnetic resonance imaging (MRI) is an important brain imaging technique that provides high-resolution information about the structure, composition, and function of the brain (3). There are several MRI modalities including structural MRI, functional MRI and diffusion MRI. Since functional connectivity of the brain is altered in ASD, this study uses resting-state functional MRI to automatically diagnose ASD (7).

Resting-state functional connectivity refers to the temporal correlation of neuronal activity between distinct brain regions when the brain is at rest, which means the individual is not doing a specific task. It reveals insights into the brain’s inherent organization and how different regions communicate with one another. This method is especially valuable for researching neurodevelopmental disorders such as ASD, because abnormalities in functional connectivity might show impaired brain networks associated with the disorder.

In recent years, methods based on machine learning have been widely used to make distinctions between people with ASD and healthy controls (HCs) by studying brain connections. Several algorithms have been employed for this purpose, including support vector machine (SVM), random forest (RF), K-nearest neighbor (KNN), and artificial neural network (ANN). These algorithms have shown good results in identifying ASD using neuroimaging data. SVM is commonly employed because to its success in high-dimensional datasets, but RF offers robust classification via ensemble learning. KNN is advantageous due to its simplicity and effectiveness in small datasets, whereas ANN, with its deep learning capabilities, has demonstrated higher performance in spotting complex patterns in brain circuitry. Given their effective application in earlier research, these four algorithms were selected for this investigation.

Several studies have explored the use of machine learning techniques to differentiate between individuals with autism and HCs based on differences in brain connections. Iidaka et al. used resting-state functional magnetic resonance imaging (fMRI) and a neural network in a study. The correlation matrix calculated from the resting-state fMRI which indicated functional connections between different brain regions was entered into a probabilistic neural network (PNN) for classification. This study achieved an accuracy of approximately 90% in classifying the two groups (8). Mahanand et al. also successfully detected ASD by identifying differences in brain activity using Resting-state fMRI. Their approach was validated using the ABIDE database, and they achieved accuracies of 78.6% for adolescent men, 85.4% for adult men, 86.7% for adolescent women, and 95% for adult women using SVM classification (9). Similarly, Sadeghi et al. proposed an automatic screening method for recognition of ASD from HC based on their brain functional abnormalities. In this paradigm, brain functional networks of 60 adolescent and young adult males (29 ASDs and 31 HCs) were estimated from subjects’ task-free fMRI data. Then, autism screening was developed based on characteristics of the functional networks using graph theory. Performance of the system was verified using various classification techniques. The SVM showed superiority to others with an accuracy of 92% (4). In another study by Bi et al. autistic individuals were classified from healthy individuals using a randomized vector classification machine and multiple SVMs. They proposed a randomized SVM clustering method to identify control and autistic individuals, achieving an accuracy of 96.15% using the optimal feature set (10). Fredo et al. classified autism and control individuals using resting-state with conditional random and random tree algorithms. They utilized fMRI images from ABIDE-I and ABIDE-II databases and achieved a classification accuracy of 65% using the RF classifier (11). Furthermore, Kazeminejad et al. conducted a study involving 816 subjects from the ABIDE database, utilizing graph theory for resting-state analysis. The SVM classification model was employed to differentiate between autistic and control subjects yielding an accurate classification model with 95% accuracy, 97% sensitivity, and 95% specificity (12). Lastly, Thomas et al. employed three-dimensional convolutional neural networks (3D-CNN) on resting-state data to classify the ASD group. The study utilized data from the ABIDE-I and ABIDE-II databases and achieved an accuracy of 66% (13).

While previous research focused on the use of machine learning algorithms for ASD diagnosis, this work focuses on the use of resting-state fMRI and advanced machine learning algorithms to identify particular brain areas linked with ASD in children. Our research intends to fill a gap in the literature by developing a more accurate and automated technique for detecting ASD, which might help existing clinical examinations.

Recent studies have explored similar graph-based approaches in neurodevelopmental disorders. For instance, Zhao et al. in 2022 applied a dynamic connectivity approach to ADHD classification, demonstrating the potential of temporal graph models (14). Another study by Tong et al., in 2024 examined EEG-based network connectivity patterns in ASD, revealing associations with symptom severity (15). Comparing our results with these findings highlights the broader applicability of graph-based methods across neurodevelopmental conditions and underscores the need for multimodal approaches.

The purpose of this study is to create an automated method for diagnosis ASD in children using resting-state fMRI and machine learning techniques. We believe that certain patterns of functional connectivity in the frontal, temporal, and cingulum areas can accurately distinguish between children with ASD and HCs. Firstly, the resting-state images were preprocessed from the ABIDE database. Then, a brain network was created for each subject which represent the functional connections between different brain regions. After that, a brain network graph was created. Because the aim was to analyze all brain areas, the graph method was used. The optimal features were given to four model prediction algorithms for diagnosing autism. The proposed method was able to distinguish the autism group from the control group with high accuracy.

Methods

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The steps outlined in Figure 1 illustrate the methodology employed in this scientific article.

Figure 1 Proposed method for automatic diagnosis of autism spectrum disorders in children through resting-state fMRI with machine vision. fMRI, functional magnetic resonance image.

Participant selection

In this study, structural and resting-state functional magnetic resonance images from the Autism Brain Imaging Data Exchange (ABIDE) site were used (https://fcon_1000.projects.nitrc.org). Specifically, we used images from the Langone Medical Center, NYU: site 1 model (NYU). The dataset consisted of brain images of 78 children. The NYU site was chosen because it provided high-quality imaging data and had a significant number of pediatric individuals. However, we recognize that utilizing data from a single site may restrict the generalizability of our findings. Future research should integrate data from many sites in the ABIDE dataset to test the suggested approach across populations.

As manual quality control (QC) was not yet available in ABIDE-II, we performed automated QC by selecting subjects who retained at least 100 frames or 4 min of fMRI scans after motion scrubbing. Motion scrubbing was performed based on framewise displacement (FD), discarding one volume before and two volumes after a frame if the FD was greater than 0.5 mm (4).

The initial dataset of 78 cases was reduced to 52 in order to focus on children aged 5–10 years, as early detection is critical for effective treatment. This age range was chosen to guarantee that the suggested technique is appropriate for the pediatric population, because early detection has the greatest influence on treatment outcomes. The sample size comprised 26 individuals diagnosed with ASD and an equal number of HC subjects. There were no significant differences in age and gender between ASD and HC subjects. Consequently, they were not included as covariates. Participant demographics are shown in the Table 1.

MRI acquisition

All MRI data were acquired on a Siemens 3 Tesla scanner with Allegra model. The output of the scanner was in Neuroimaging Information Technology Initiative (NIFTI) format, and the Magnetic field strength was 3 tesla.

The anatomical images were acquired using a 3D TFL sequence with a repetition time (TR) of 3.25 ms and an echo time (TE) of 3.25 ms. The field of view was 256×256 mm², and the slice thickness was 1.33 mm, with 128 slices acquired. The acquisition matrix was 256×192, and the total acquisition time was 8 minutes and 7 seconds.

For the resting-state fMRI, a standard EPI sequence was used with a TR of 2,000 ms and a TE of 30 ms. The field of view was 240×240 mm², with a slice thickness of 3 mm. A total of 34 slices were acquired, with an acquisition matrix of 80×80, and the total acquisition time was 6 minutes.

Preprocessing

Fifty-two structural and resting-state functional images of autistic and HC subjects were preprocessed for farther analysis. The preprocessing way for the sMRI and rest-state fMRI images were performed using DPARSF (data processing adjunct for resting-state) (http//rfmri.org/DPARSF) toolbox in MATLAB.

sMRI preprocessing steps was include: skull stripping (removing non-brain tissues), segmentation (gray matter, white matter, and cerebrospinal fluid), and spatial normalization (register the structural images to the MNI space). Also preprocessing steps of resting-state was slice timing correction, motion correction (realign all volumes), temporal filtering (bandpass filtering), skull stripping, and spatial normalization (register the functional images to the MNI space).

The next key step is to register resting-state functional images to the structural (anatomical) images. In DPARSF, coregistration between functional and structural images is performed automatically during the preprocessing pipeline. This process ensures that the fMRI data and anatomical data are spatially aligned within the same coordinate system (29).

Brain network

In this study, we obtained a connectome or functional connectivity matrix to detect ASD, because functional brain connections are changed in ASD. This matrix reveals the correlation between the mean values of time series obtained from the region of interest (ROI). Each cell of the matrix contains a Pearson correlation coefficient (PCC), with each row representing an ROI. The PCC range from −1 to 1, where a value of 1 indicates a strong correlation between the two brain regions and vice versa (16). The elements along the main diagonal of the matrix are all equal to 1 as a signal is perfectly consistent with itself in terms of correlation. Moreover, based on the exchange property of calculating the correlation coefficient, the matrix is symmetric. Since the automated anatomical labelling (AAL) atlas has 116 areas, so the dimensions of each of the connectivity matrices are 116×116. An example of such a matrix for a control participant is shown in Figure 2A.

Figure 2 Brain network and graph. (A) Brain network of AAL atlas regions. (B) A graph made of the whole brain (nodes are shown with blue spheres and connections with yellow color). AAL, automated anatomical labeling.

Graph

There are various methods for analyzing resting-state images, including seed-based, independent component analysis (ICA), graph, clustering algorithms, neural networks, and pattern classification. Since the aim of the present study is to analyze rest-state functional connectivity across all brain regions, the graph method is used. Each graph network G is shown as G = (N, K), where N represents the number of nodes, encapsulating the brain regions under study, while K signifies the number of edges in the graph, symbolizing the connections between these nodes. Using the brain network and the constructed connection matrix, graph analysis can be performed by using BRAGH software (17). The graph made of the whole brain is shown in Figure 2B.

Feature extraction

The process of extracting features from Resting-state functional magnetic resonance images involves two main steps:

• Extraction of Pearson coefficients from the functional connectivity matrix: this feature is used to investigate the decrease or increase of functional connections between each pair of the AAL Atlas brain region. PCC between each pair of fisher functional connections matrix extracted.
• Extraction of graph analysis parameters: in this step, the node and global parameters of a non-directional weight graph made from the whole brain are extracted by Braph software. The node parameters extracted from the non-directional weight graph of the whole brain can be found in Table 2, whereas the global parameters extracted from the same graph are listed in Table 3. A feature matrix is created for each autism and control group based on the extracted features. Since correlation matrices of functional connections are symmetric, only the upper triangle of these matrices is considered as a feature, while the remaining elements are considered additional. The total number of elements in each brain network matrix is 13,456. Among these, 116 elements correspond to self-connections (diagonal elements) and are excluded. Consequently, 13,340 features remain, and half of these (6,670) are extracted and converted into a one-dimensional vector. Additionally, the nodal and global parameters of the undirected weighted graph made of the whole brain are added to this vector to form the feature matrix of Resting-state functional magnetic resonance images. After data preprocessing, a feature matrix with dimensions of 26×7,144 is obtained; which is saved for further processing. The total number of features extracted from resting-state functional magnetic resonance images was 6,894, and each group included 26 subjects. The dimensions of the feature matrices related to resting-state images are shown in the Table 4.

Feature selection

To remove irrelevant features before classification and improve machine vision accuracy, we use the fast correlation-based filter (FCBF) method in Python software to select basic features for separation between two groups based on Khadem-Reza et al.’s research (3). FCBF is a multivariate feature selection method that considers class relatedness and dependencies between each feature pair. Based on information theory, FCBF uses symmetric uncertainties to compute feature dependencies and class relevance (18).

Classification

Automated classification using structural MRI is performed using four machine vision algorithms including SVM, RF, nearest neighbour, and ANN. The implementation of these algorithms is done using Python 3.8.3 software with 10-fold cross-validation (K fold =10). A Multi-Layer Perceptron (MLP) with three hidden layers, ReLU activation functions, and an Adam optimizer is the ANN used in this investigation. Based on cross-validation results, k=3 was chosen for KNN and a radial basis function (RBF) kernel was utilized for SVM. The output of the classification algorithms is a confusion matrix that can be used to assess the algorithms. Finally, the performance of all four classifications is assessed and the system’s best intergroup classification performance is determined.

Results

Feature extraction

Selected features of resting-state functional magnetic resonance images are given in the Table 5. To assess which brain regions exhibit significant functional differences between the two groups, the features in Table 5 are classified according to their respective regions. This analysis is illustrated in Figure 3.

Figure 3 Repetition number of brain regions in selecting features.

Classification

In all the mentioned classification steps, the output of the classification algorithms is a convolution matrix, which contains the following parameters: true positive (TP), true negative (TN), false positives (FP), and false negative (FN).

The evaluation of classification systems using the mentioned indicators is done by calculating the following parameters:

Specificity: represents the proportion of negative results that have been correctly diagnosed, for example, the percentage of people who, according to the model’s prediction, do not have ASD and in fact do not exhibit any symptoms of this disorder.

Recall or sensitivity: represents the proportion of positive results that are correctly diagnosed, for example, the percentage of people who are diagnosed with ASD according to the model’s prediction and indeed have this disorder.

Accuracy: denotes the degree of proximity between the measured value and the actual value. In order to achieve accuracy, precision must be high, but the opposite is not necessarily true. A high degree of bias and variance is indicative of low accuracy.

Precision: demonstrates how closely consecutive measurements of a particular value align with each other. How to calculate these parameters is shown in the following equations.

Specificity=TNFP+TN×100%[1]

Sensitivity=TPTP+FN×100%[2]

Precision=TPTP+FP×100%[3]

Accuracy=TP+TNTP+TN+FP+FN×100%[4]

The components of the confusion matrix resulting from each of the classification algorithms are given in the Table 6. Also, the evaluation parameters for each classification system are calculated and displayed in the Table 7. According to the obtained results, the ANN automatic classifier has the best performance compared to other algorithms for classifying the autism group from the control group, and with 90.38% accuracy, it has the ability to distinguish these two groups.

Discussion

The purpose of the current study is to identify ASD automatically using features extracted from resting-state fMRI images. The extracted features with maximum difference between two groups are selected by FCBF algorithm. These features are shown in the Table 6. As shown in Figure 3, the frontal region was present in 19 of the selected features, and this region is known to have the highest functional abnormality in the autism group in this study. Subsequently, the comparison between the two groups revealed notable differences in the temporal and cingulum regions. The frontal lobes are important for voluntary movement, expressive language and for managing higher level executive functions. Executive functions refer to a collection of cognitive skills including the social skills, capacity for planning, organizing, initiating, self-monitoring and controlling one’s responses in order to achieve a specific goal (19). Moreover, the main functions of the temporal lobes are understanding language, memory acquisition, face recognition, object recognition, perception, as well as auditory information processing (5). Considering that autism disorder is a disorder of cognitive skills and social, perception, recognition and language skills are also damaged in this disorder (20), it is quite expected that the frontal and temporal brain areas in autistic people is identified with the most functional impairment and this indicates that the result of the present study is confirmed. Numerous studies have shown that individuals with autism have functional impairments in the frontal and temporal lobes of the brain. For example, one study illustrated a reduced connectivity between the frontal and temporal lobes in individuals with autism, which is critical for cognitive and thinking processes. This decreased connectivity may contribute to difficulties in processing information and performing complex cognitive tasks (21). Studies have also shown that individuals with autism have decreased activity in the temporal lobe, which is involved in cognitive and social processes. This decreased activity may contribute to difficulties in processing social information and interpreting emotions (22). Based on another study, it can be concluded that the frontal and temporal lobes of the brain have functional impairments in individuals with autism. These impairments may manifest as symptoms and signs of autism. Also, the cingulum is the major fiber system connecting the cingulate and surrounding medial cortex and medial temporal lobe internally and with other brain areas. It is important for social and emotional functions related to core symptomatology in ASDs (23).

We conducted a comparison of four popular machine-learning classifiers, namely SVM, RF, K-nearest neighbour (KNN) algorithm and ANN in order to evaluate their performance in generating diagnostic models for ASD. Our findings indicate that ANN superior to the other three classifiers. According to Table 7 in the present study, SVM, RF, KNN and ANN achieved accuracy of 88.46%, 73.07%, 82.69%, 90.38% respectively in the classification and diagnosis of ASD using resting-state fMRI. We conclude that ANN is the best approach for neuroimaging data mining with small sample sizes. ANNs can be advantageous in small sample scenarios because of their ability to model complex relationships, utilize regularization techniques, leverage pre-trained models (transfer learning), and learn non-linear patterns (21).

Our study aimed to establish baseline performance using conventional machine learning techniques, even though graph neural networks (GNNs) provide promising capabilities for studying brain connectivity. GNNs will be used in future research to investigate their potential for improving the classification of ASD.

While the process of extracting features and feeding them into machine learning algorithms is not new, our study advances the field by using these techniques to resting-state fMRI data for the diagnosis of ASD in children. This method has the potential to increase the accuracy and efficiency of ASD diagnosis, especially in pediatric populations.

While our findings are valuable, it is important to note that they are based on a smaller sample size than previous studies, such as Kazeminejad et al. (12), who achieved greater accuracy using an extensive dataset. Future study should try to test our findings on larger, more diverse datasets that will guarantee the suggested method’s resilience.

The proposed method has potential applications in clinical settings, particularly for aiding early diagnosis and monitoring of ASD. But in practice, a number of factors need to be carefully taken into account. Initially, modifications might be required to customize the model for various patient demographics, guaranteeing that it extends beyond the particular dataset employed in this investigation. Second, large-scale validation is needed across diverse cohorts to confirm its reliability. Finally, integrating this approach into existing clinical workflows would require collaboration with healthcare professionals to optimize usability and interpretability.

In addition to the graph-theoretic measures employed in this study, other network properties such as modularity and participation coefficient could offer further insights into ASD-related connectivity differences. Modularity, which quantifies the strength of community structure in a network, may help identify disruptions in functional segregation. The participation coefficient, on the other hand, characterizes how brain regions interact across different functional modules, potentially highlighting atypical hub connectivity in ASD. Future studies should explore these measures to enhance the characterization of ASD-related network abnormalities.

Several confounding factors may influence both fMRI data quality and the classification outcomes in this study. Comorbidities such as ADHD, anxiety, and epilepsy are commonly observed in individuals with ASD and could introduce variability in functional connectivity patterns. Additionally, the use of psychotropic medications may alter brain activity and should be accounted for in future analyses. Socioeconomic factors, including access to healthcare and educational background, may also play an indirect role in shaping neural and behavioral outcomes. Addressing these variables in future research through stratified analyses or covariate adjustments could enhance the robustness of findings.

The AAL atlas was used in this study to parcellate the brain, but other atlases, like the Schaefer and Yeo atlases, might offer more precise spatial resolution and complementary information. The specificity of network analyses may be increased by the Schaefer atlas, which uses gradient-based cortical parcellation to capture more functionally homogeneous regions. In a similar vein, a more physiologically significant depiction of extensive brain networks may be obtained through the Yeo atlas, which is predicated on resting-state functional connectivity. Subsequent investigations may examine various parcellation techniques to evaluate their influence on classification efficacy and outcomes of network analysis.

Limitation and future work

Several limitations should be acknowledged, despite the fact that the proposed method shows promising classification performance. First, in order to verify generalizability, external validation on separate cohorts is required, as the study was carried out on a particular dataset. Second, combining multimodal imaging data (diffusion MRI, EEG, etc.) may offer a more thorough comprehension of network changes linked to ASD. Third, there is still work to be done to improve the clinical interpretability and transparency of graph-based models, as they remain difficult to interpret. To increase classification accuracy and gain more neuroscientific understanding, future research should work to improve these techniques, incorporate multimodal data, and investigate deep learning-based graph approaches.

Conclusions

The proposed framework was tested on 52 subjects from the ABIDE data set consisting of 26 patients with ASD and 26 HC subjects. Since clinical evaluations for the diagnosis of autism are extremely time-consuming and depend on the expertise of a specialist, the importance of intelligent diagnosis of this disorder becomes clear.

In conclusion, this study proposed an innovative approach to diagnosing autism by analyzing localized brain areas. The results demonstrated the high accuracy of the proposed framework and identified specific brain regions that can be further studied for better understanding and treatment of autism. The development and use of intelligent systems for diagnosing autism have the potential to greatly improve the efficiency and reliability of the diagnostic process.

Furthermore, this study identified the frontal, temporal and the cingulum as the regions with the highest differentiation between autistic and control children. Understanding these brain areas can provide insights into the neural underpinnings of the disorder and aid in clinical diagnoses.

Acknowledgments

This paper was extracted from an MS.c thesis of Medical Physics by the author Zahra Khandan Khadem-Reza.

Footnote

Funding: This work was supported by the Research Deputy of MUMS for financial support of this project, numbered (980858) (ethics code: IR.MUMS.MEDICAL.REC.1398.717).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1402/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Faraj R, Zare H. Evaluation of white matter tracts in autistic individuals: a review of diffusion tensor imaging studies. SSU Journals 2021;29:3539-55.
2. Khadem-Reza ZK, Zare H. Evaluation of brain structure abnormalities in children with autism spectrum disorder (ASD) using structural magnetic resonance imaging. The Egyptian Journal of Neurology, Psychiatry and Neurosurgery 2022;58:1-14.
3. Khadem-Reza ZK, Zare H. Automatic detection of autism spectrum disorder (ASD) in children using structural magnetic resonance imaging with machine vision system. Middle East Current Psychiatry 2022;29:54.
4. Sadeghi M, Khosrowabadi R, Bakouie F, Mahdavi H, Eslahchi C, Pouretemad H. Screening of autism based on task-free fMRI using graph theoretical approach. Psychiatry Res Neuroimaging 2017;263:48-56. [Crossref] [PubMed]
5. Yu Q, He Z. Comprehensive investigation of temporal and autism-associated cell type composition-dependent and independent gene expression changes in human brains. Sci Rep 2017;7:4121. [Crossref] [PubMed]
6. Jones RM, Lord C. Diagnosing autism in neurobiological research studies. Behav Brain Res 2013;251:113-24. [Crossref] [PubMed]
7. Khandan Khadem-Reza Z, Shahram MA, Zare H. Altered resting-state functional connectivity of the brain in children with autism spectrum disorder. Radiol Phys Technol 2023;16:284-91. [Crossref] [PubMed]
8. Iidaka T. Resting state functional magnetic resonance imaging and neural network classified autism and control. Cortex 2015;63:55-67. [Crossref] [PubMed]
9. Mahanand B, Vigneshwaran S, Suresh S, Sundararajan N. editors. An enhanced effect-size thresholding method for the diagnosis of Autism spectrum disorder using resting state functional MRI. 2016 Second International Conference on Cognitive Computing and Information Processing (CCIP); 2016: IEEE.
10. Bi XA, Wang Y, Shu Q, Sun Q, Xu Q. Classification of Autism Spectrum Disorder Using Random Support Vector Machine Cluster. Front Genet 2018;9:18. [Crossref] [PubMed]
11. Agastinose Ronickom JF, Jahedi A, Reiter M, et al. Diagnostic Classification of Autism using Resting-State fMRI Data and Conditional Random Forest. Conference proceedings: ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference 2018:1148-51.
12. Kazeminejad A, Sotero RC. Topological Properties of Resting-State fMRI Functional Networks Improve Machine Learning-Based Autism Classification. Front Neurosci 2018;12:1018. [Crossref] [PubMed]
13. Thomas RM, Gallo S, Cerliani L, Zhutovsky P, El-Gazzar A, van Wingen G. Classifying Autism Spectrum Disorder Using the Temporal Statistics of Resting-State Functional MRI Data With 3D Convolutional Neural Networks. Front Psychiatry 2020;11:440. [Crossref] [PubMed]
14. Zhao K, Duka B, Xie H, Oathes DJ, Calhoun V, Zhang Y. A dynamic graph convolutional neural network framework reveals new insights into connectome dysfunctions in ADHD. Neuroimage 2022;246:118774. [Crossref] [PubMed]
15. Tong X, Xie H, Fonzo GA, Zhao K, Satterthwaite TD, Carlisle NB, Zhang Y. Symptom dimensions of resting-state electroencephalographic functional connectivity in autism. Nat Ment Health 2024;2:287-98. [Crossref] [PubMed]
16. Cohen I, Huang Y, Chen J, Benesty J, Benesty J, Chen J, et al. Pearson correlation coefficient. Noise Reduction in Speech Processing 2009:1-4.
17. Ruiz PEG, Volpe G, Pereira JB, Garcia AC, Mijalkov M, Chang YWG. editors. BRAPH 2.0: Software for the analysis of multilayer brain connectivity with graph theory and machine learning. Emerging Topics in Artificial Intelligence (ETAI) 2021;2021:SPIE.
18. Senliol B, Gulgezen G, Yu L, Cataltepe Z. editors. Fast Correlation Based Filter (FCBF) with a different search strategy. 2008 23rd International Symposium on Computer and Information Sciences; 2008:IEEE.
19. Chayer C, Freedman M. Frontal lobe functions. Curr Neurol Neurosci Rep 2001;1:547-52. [Crossref] [PubMed]
20. Mohammad-Rezazadeh I, Frohlich J, Loo SK, Jeste SS. Brain connectivity in autism spectrum disorder. Curr Opin Neurol 2016;29:137-47. [Crossref] [PubMed]
21. Feczko E, Miezin FM, Constantino JN, Schlaggar BL, Petersen SE, Pruett JR Jr. The hemodynamic response in children with Simplex Autism. Dev Cogn Neurosci 2012;2:396-408. [Crossref] [PubMed]
22. Schroeder JH, Desrocher M, Bebko JM, Cappadocia MC. The neurobiology of autism: Theoretical applications. Research in Autism Spectrum Disorders. 2010;4:555-64.
23. Hau J, Aljawad S, Baggett N, Fishman I, Carper RA, Müller RA. The cingulum and cingulate U‐fibers in children and adolescents with autism spectrum disorders. Human Brain Mapping 2019;40:3153-64. [Crossref] [PubMed]