1 Introduction

Autism Spectrum Disorder (ASD) identification has emerged as a crucial issue within the field of neurodevelopmental research, as the prevalence of ASD grows worldwide. ASD is a complex and diverse neurodevelopmental syndrome marked by difficulties in social communication and repetitive activities. Identifying ASD early in a child’s life is critical for appropriate intervention and support, yet it remains a significant difficulty owing to the complex and frequently subtle nature of the symptoms. Traditional diagnostic procedures depend mainly on clinical observations, behavioral evaluations, and interviews, which are subjective, time-consuming, and may lack the sensitivity required for early diagnosis (1). In recent years, the use of deep learning approaches into ASD detection has showed great potential and drew major interest from researchers and clinicians. Deep learning, a form of machine learning that uses artificial neural networks modelled after the human brain, provides a unique technique to identify the detailed patterns and traits linked with ASD in neuroimaging data. Magnetic resonance imaging (MRI) and functional MRI (fMRI) have emerged as critical methods for understanding the brain’s anatomical and functional connections in a non-invasive manner (2). Deep learning methods, particularly neural networks with numerous layers, have shown the ability to automatically learn and extract nuanced patterns from neuroimaging data, allowing for the detection of minor anomalies linked with ASD. Minor anomalies are minor differences in the structure and function of the brain, undetectable using conventional methods and often too minor to be easily visible. Such minor differences include slight variations in the amygdala and cerebellum, abnormal connectivity patterns between areas of the brain, alterations in cortical thickness or volume, or microstructural changes in white matter. Advanced deep learning algorithms identify such minor differences when processing neuroimaging data. Thus, these minor differences are associated with some cognitive and behavioral features of ASD, which allows for earlier and more precise diagnosis. Deep learning has the ability to overcome various issues that standard approaches confront while identifying ASD. Deep learning algorithms can effectively handle large volumes of data, identifying nuanced patterns that would be difficult to detect manually. Furthermore, these models may incorporate multimodal information by incorporating data from a variety of sources, including structural and functional MRI, genetic markers, and behavioral evaluations. The end result is a more thorough and nuanced knowledge of ASD’s neurological roots, which might lead to more accurate and early detection. Furthermore, the incorporation of attention processes into deep learning models enables the prioritizing of key variables, improving interpretability and enabling the discovery of biomarkers linked with ASD. Understanding how deep learning models make decisions is crucial for bridging the gap between their complex, ‘black box’ nature and the clinical insights needed to make informed decisions in practice As deep learning advances, new architectures, optimization approaches, and hybrid models are being developed to improve the accuracy, resilience, and generalizability of ASD diagnosis models.

2 Related work2.1 Recognizing ASD via various deep learning models

Functional Connectivity in ASD (3) is complex and shows hypo- and hyper-connectivity within many brain networks. Dimensional clustering can determine unique FC subtypes that relate to specific behavioral characteristics, hence unearthing a wide variety of connections between brain activity and behavior. Recognition of common patterns in FC would explain the heterogeneity of ASD better, facilitate better diagnosis, and develop tailored interventions that bridge neural changes to observable symptoms. The use of neuroimaging methods to better understand and diagnose neurodevelopmental disorders suggests a possible option for enhancing clinical practice. A technique for estimating effective connectivity in brain networks (BNs) (4) using EEG data, with a focus on children with attention-deficit hyperactivity disorder (ADHD). Their findings revealed substantial differences in directed information transmission across EEG electrodes in ADHD patients compared to healthy controls, with discriminative power notably high in the theta-band, which is associated with focus and motor activity. Similarly (5), used functional MRI (fMRI) data from the Autism Brain Imaging Data Exchange (ABIDE) dataset to solve the problem of multi-site data aggregation in autism diagnosis. They used data harmonization approaches to improve classification accuracy and discovered insights into ASD pathogenesis through network analysis. With machine learning techniques applied to fMRI functional connectivity, promising detection of ASD can be automated. The critical advances include integration of temporal dynamics, multiscale data, and focused analysis of brain networks relevant to ASD. Even though performance keeps improving, more work is needed to create robust and interpretable models for clinical application. Combining data modalities with larger datasets may further propel this field toward reliable computer-aided diagnosis of ASD (6). Their research produced encouraging results and revealed changes in brain network structure related to ASD. Furthermore (7), the significance of using sequential information from task-based fMRI for synthetic data augmentation in ASD diagnosis is demonstrated, demonstrating the potential for improving diagnostic accuracy and comprehending the underlying illness. Together, these studies highlight the value of neuroimaging techniques and advanced machine learning approaches in understanding the intricacies of neurodevelopmental disorders, opening the path for more effective diagnostic tools and deeper insights into the brain.

An Unsupervised Contrastive Graph Learning (UCGL) framework for resting-state fMRI research (8) underlines the difficulty of acquiring labelled training data in clinical practice and suggests a pretext model trained on unlabeled data for subsequent illness diagnosis tasks. The UCGL framework is tested on three rs-fMRI datasets, and it outperforms current techniques in the automated diagnosis of major depressive illness, ASD, and Alzheimer’s disease. ASD is diagnosed using structural (sMRI) and functional (fMRI) MRI modalities (9). To solve the low data availability, transfer learning is used in conjunction with four vision transformers and a 3D-CNN model. The investigations use several ways to generate data and extract slices from raw 3D sMRI and 4D fMRI images, yielding ground breaking findings. The brain disease categorization in resting-state functional magnetic resonance imaging (rs-fMRI) data study (10). Using multi-omics data, the team created eleven networks that depicted various facets of the brain. Kullback-Leibler divergence is used in their methodology, Graph convolution techniques are then applied to learn gene-disease connections. Functional magnetic resonance imaging (fMRI) data has been more important in neuroscience research recently for the identification of neurological illnesses and for the comprehension of cognitive processes. Many research has looked at various approaches to deal with issues such noise, limited sample numbers, and the requirement for interpretability that arise while processing fMRI data. Applying fMRI data to categorize neurological disorders or cognitive function has been the subject of several research. Using fMRI-derived brain graphs, GroupINN (11)—a grouping-based interpretable neural network—classifies cognitive performance well. Node grouping is included into the architecture of this model, which jointly learns these groups and extracts graph properties. Using resting-state fMRI time-series data, networks of long- and short-term memories (LSTMs) (12) can be used to directly diagnose autism spectrum disorders (ASD). They used the extensive, multi-site Autism Brain Imagery Data Exchange (ABIDE) I database for testing and training, and as a result, their classification accuracy was higher than that of earlier techniques. By applying algorithms that use deep learning to the ABIDE dataset (13), it was possible to identify ASD patients only by looking at patterns of brain activation. This approach achieved a remarkable degree of accuracy and revealed functional connectivity patterns linked to ASD. ASD-DiagNet (14), a system for ASD identification using fMRI data, was developed utilizing the 1,035 participants from 17 imaging facilities in the Autism Brain Imaging Information Exchange dataset. Techniques include a hybrid strategy for feature extraction that uses a single layer the perceptron and an autoencoder, as well as a linear interpolation-based data augmentation technique. In order to overcome the shortcomings of conventional behavioral observation-based diagnostic techniques, the research emphasizes the importance of developing machine learning infrastructures for qualitative diagnosis of varied mental disorders like autism spectrum disorder.

In these studies, the ABIDE dataset has become a notable resource offering a big, multi-site collection of fMRI data for research on ASD. By using this dataset, researchers may test and train their models on a variety of samples, which improves the generalizability of their conclusions. This research uses a range of methodologies, from deep learning architectures to conventional machine learning approaches. While deep learning models frequently give better performance but lack transparency, classical models could offer interpretability. Thus, by putting forth models which are both practical and understandable, recent initiatives have sought to close this gap.

2.2 GNN-based ASD recognition

Adversarial Graph Contrastive Learning (A-GCL) (15) is a method for identifying neurodevelopmental problems using fMRI data. The model employs a graph neural network (GNN) based on graph contrastive learning, with graphs generated from fMRI data. A-GCL outperforms three datasets, Autism Brain Imaging Data Exchange (ABIDE) I, ABIDE II, and attention deficit hyperactivity disorder (ADHD), over three atlases. While an A-GCL model may provide a good theoretical fit, this may be underpinned with several limitations including high-quality data in fMRI, inability to generalize across very diverse populations, low interpretability, and computing demands. Specifically, the atlases used need to be determined and only the resting-state fMRI is adopted. A Multi-Scale Dynamic Graph Learning (MDGL) (16) framework for detecting brain disorders utilizing resting-state functional magnetic resonance imaging (rs-fMRI) data. They use various brain atlases to build multi-scale dynamic functional connectivity networks (FCNs) and graph neural networks to extract spatiotemporal information from them. ASD categorization based on brain functional activity and gene expression using an attention-based graph neural network (GNN) (17). Their findings highlighted the significance of customized information in ASD diagnosis and biomarker identification. By combining individual brain topology and graph data. Furthermore, they identified brain regions important for ASD, such as the social-brain circuit and default-mode network, and discovered ASD-related genes using functional MRI data and gene expression analysis, demonstrating the potential of their approach for effective ASD diagnosis and biomarker identification. Present the Autism Spectrum Disorder-based Attention GNN and Crossover Boosted Meerkat Optimization (ASD-AttGCBMO) (18) algorithm. The suggested technique uses structural Magnetic Resonance Imaging (sMRI) data from the ABIDE 1 dataset for preprocessing to improve picture quality. Surface-based analysis and voxel-based morphometry (VBM) both extract significant information such surface area, cortical thickness, shape descriptors, and brain volumes. To address issues such as overfitting and class imbalance, the model utilizes attention GNNs with crossover-boosted meerkat optimization. PLSNet (19) is a position-aware graph-convolution-network-based model for ASD diagnosis that uses functional MRI (fMRI) data. PLSNet includes a time-series encoder for feature extraction and a connection generator to represent long-term relationships. PLSNet includes a time-series encoder for feature extraction and a connection generator to represent long-term relationships. Position embedding and a rarefying approach are used to solve challenges such as brain region variation and dimensionality complexity. The work delivers cutting-edge performance on the Autism Brain Imaging Data Exchange dataset.

The neuroimaging methods, namely magnetic resonance imaging (MRI) and functional MRI (fMRI), to diagnose Autism Spectrum Disorder (ASD). Each research suggests a unique methodology, such as machine learning techniques, generative adversarial networks (GANs), unsupervised contrasting graph learning, and adversarial self-supervised GNNs. A Conditional Generative Adversarial Network (cGAN) for predicting ASD (20). The researchers emphasize the limits of classic supervised machine learning techniques when dealing with tiny datasets, and they provide a cGAN that surpasses normal GANs in terms of prediction accuracy. A graph attention network (GAT) (21), based on spatially restricted sparse functional brain networks (FBNs), was used to diagnose ASD. They developed a unique approach, Pearson’s correlation-based Spatial Constraints Representation (PSCR), for estimating FBN structures and feeding them into a GAT for classification. Their trials using the ABIDE I dataset demonstrated the superiority of the PSCR technique as well as the influence of various FBNs on classification outcomes. Their suggested system produced encouraging classification results, surpassing rival approaches and offering insights for future illness detection using FBN and GNN frameworks.

The focus is on ASD, and task-based functional magnetic resonance imaging (fMRI) data is used. The authors focus on data-driven learning algorithms for biomarker identification and outcome prediction. Their deep learning methods use GNNs (GNNs) (22) for spatial variables and Long Short-Term Memory (LSTM) networks for temporal features. The chapter emphasizes the significance of dynamic connectivity changes and provides a more comprehensive, integrated model that includes spatiotemporal aspects as well as neural ordinary differential equations. ASD (23), utilizing a thorough examination of multi-modal imaging markers. Their dual-branch GNN performs a major diagnosis by extracting and combining data from structural and functional magnetic resonance imaging. The study also uses a perturbation model to find brain imaging signals and a neuro-transcriptomic joint analysis to reveal putative genetic biomarkers related with ASD brain development. Contribute to psychiatric diagnosis (24) using brain-networks by presenting a Granger causality-inspired GNN (CI-GNN). The model strives for interpretability without resorting to post-hoc interpretative paradigms.

A graph neural network architecture called BrainGNN (25) was used to analyze fMRI datasets from the Human Connectome Program (HCP) 900 Participant Release and Autism Spectrum Disorder (ASD). Their solution leverages both functional and topological knowledge gathered from fMRI data by using new ROI-aware graph convolution layers (Ra-GConv). Furthermore, BrainGNN uses ROI-selection layer pooling (R-pool) to emphasize important brain areas, making it easier to read. Regularization terms are suggested to promote flexible modeling of single or group-level patterns and fair ROI selection, such as units loss, topK pooled (TPK), which loss, and group-level constancy (GLC) loss.

2.3 GCN-based ASD recognition

In recent research, there has been an increasing emphasis on using functional brain networks (FBN) to classify neurological illnesses, particularly Autism Spectrum Disorder (ASD). MVS-GCN (2), a multi-view graph convolution network guided by previous brain structure learning, to overcome the issues presented by subject heterogeneity and noise correlations in brain networks. Their machine learning technique not only helps to classify neurological illnesses, but also gives an interpretable framework for deeper insights into the brain network. A joint learning architecture of multi-level dynamic brain networks for the diagnosis of ASD (26). They overcome the constraints of previous graph convolutional network (GCN)-based techniques by allowing bidirectional information sharing across brain networks and adding edge weight information via an edge self-attention mechanism. It allows for information complementarity across different layers of brain networks. The Autism Spectrum Disorder-based Attention Graph Neural Network and Crossover Boosted Meerkat Optimization (ASD-AttGCBMO) method (27) detects ASD using structural MRI data. Their approach uses attention graph neural networks and a crossover boosted optimization strategy to improve feature categorization between ASD and control participants. Graph Neural Network (GNN) topologies and machine learning models for analyzing rs-fMRI data to better understand schizophrenia (28). They train deep graph convolutional neural networks (DGCNNs) and machine learning models with graph-theoretical measurements based on functional correlations between brain areas of interest CI-GNN finds influential subgraphs associated with choices (e.g., major depressive disorder), meeting the demand for interpretable graph neural networks and highlighting the significance of causal linkages in explainability. Moving on to neurodegenerative illnesses (29), investigates early dementia prediction using fMRI data and a Graph Convolutional Network (GCN) technique A multi-task learning strategy using a knowledge graph attention network to identify both mental and physical diseases (MPD) (30) simultaneously.

Investigating disparities in brain activity to differentiate between individuals with Autism Spectrum Disorder (ASD) and those without aids in understanding the root causes of ASD, leading to enhanced diagnosis and treatment strategies. As a result, functional connectivity (FC) analysis (31) derived from resting-state functional magnetic resonance imaging (rs-fMRI) data has emerged as a potent method for assessing and charting brain activity.

Using a supervised the siamese graph convolution neural network (s-GCN) (32) as a foundation, this approach learns a graph similarity measure with a specific emphasis on comparing brain connection networks. The model takes graph structure into account by using spectral graph convolutions, which function in the graph spectral domain. This improves results on the ABIDE database. Graph neural networks, with applications in neurology and other domains, have shown promise as a method for processing graph-structured data. During the graph representation learning process, local structural information is preserved by the use of EigenPooling (33), a pooling operation based on the graph Fourier transform. By fusing layers for pooling based on EigenPooling with conventional graph convolutional layers, they create EigenGCN, a graph neural network architecture for graph classification, and demonstrate its efficacy on six widely used benchmarks. To overcome the shortcomings of current approaches in the field of resting-state functional magnetic resonance imaging (rs-fMRI) (34) research, formulate functionally connected networks as spatiotemporal graphs. To simulate the non-stationary character of functional connectivity, they present a spatio-temporal graphing convolutional network (ST-GCN) built on brief sub sequences of BOLD time series. When it comes to predicting age and gender from BOLD data, ST-GCN performs better than standard methods. It also finds key brain areas and functional linkages that are involved in the predictions.

Functional brain networks (FBN) have attracted a lot of interest as a means of diagnosing neurological disorders including autism spectrum disorders (ASD). Accurate categorization is difficult due to noisy correlations in brain networks and considerable subject heterogeneity. MVS-GCN (2), which combines graph neural networks to get efficient end-to-end representations for brain networks. To improve classification performance and find possible functional subnetworks, this approach combines multi-view graph convolutional neural networks with previous knowledge of brain anatomy. Using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and Autism Brain Imaging Data Exchange (ABIDE) datasets, the authors assess the MVS-GCN model and show that it is more effective than current techniques. This concept motivates to generate a integrating the views in a combined form with the help of transformer and introduced a new Multi-View United Transformer Block (MVUTB). The proposed model also improves the performance of class discrimination using Graph Attention Network.

3 Background3.1 Graph attention layer on ASD

The Graph Attention Layer (GAT) is critical in identifying essential patterns and characteristics in neuroimaging data for recognizing Autism Spectrum Disorder (ASD). The application of GAT is discussed here, with equations to show how it works in the context of ASD recognition.

The GAT layer in the ASD identification task receives a collection of node characteristics reflecting neuroimaging data connected with brain areas (34). Let h= N is the number of brain regions, and F is the number of related characteristics for each area. The goal is to create a new set of node functionalities  h0=, with h^01in ℝFo as Output.

The transformation starts with a common linear transformation, parameterized by a weight matrix. W in ℝFoX F, applied to each node’s features:

A self-attention mechanism is used to produce attention coefficients (eij) that indicate the significance of characteristics from node j to node i for ASD detection. The attention coefficients are generated using a shared attentional mechanism and masked attention, focusing solely on first-order neighbors.

To introduce the graph structure, masked attention is conducted, only for the node j which is a neighbor of node i, that is taking just the neighborhood Ni of node i. To make coefficients comparable across nodes, a softmax function is applied:

aij=Softmax(exp(eij)∑k∈Niexp(eik))(3)

Finally, the new node includes h^0i is calculated as a weighted sum of the original characteristics, using attention coefficients as weights.

This method efficiently captures the significance of characteristics from surrounding brain areas in the context of ASD identification (35) by utilizing the self-attention mechanism and graph structure information. The GAT layer’s capacity to selectively aggregate information depending on attention coefficients improves the overall efficacy of ASD detection in neuroimaging data.

3.2 Transformer-based ASD recognition

Transformer-based models have gained popularity in recent neuroimaging studies for detecting Autism Spectrum Disorder (ASD) (36, 37). Transformers, which were first introduced for natural language processing, have shown amazing ability to capture subtle patterns and relationships within sequential data, making them ideal for analyzing complex brain imaging data linked with ASD.

3.2.1 Overview of transformer model

The Transformer design includes an encoder-decoder structure, however in the context of ASD identification, we concentrate on the encoder. The Transformer encoder’s main components include self-attention mechanisms (38), multi-head self-attention (MHA), and feedforward neural networks (FFN).

3.2.2 Self-attention mechanisms

The self-attention mechanism is a critical component of the Transformer model, allowing it to weigh various elements of the input sequence independently. Given an input sequence, the self-attention mechanism calculates attention ratings for each element in the sequence in comparison to all other items. The attention scores are then utilized to calculate a weighted total of the input sequence, which results in context-aware representations for each piece. The mathematical expression for the self-attention process is as follows:

Attention(Q,K,V)=Softmax(QKTdk)V (5)

The query, key, and value matrices are represented by Q, K, V, respectively, and dk denotes the dimension of the key vectors.

3.2.3 Multiple-head self-attention

To increase model expressiveness, the Transformer utilizes many attention heads in tandem. Each head gains a separate understanding of the relationships within the input sequence. The outputs from the several heads are concatenated and linearly processed to generate the final attention output. The formula for multi-head self-attention is as follows:

MHA(Q,K,V)=Concat(Head1,……,Headn)Wo(6)

Where Headi=Attention(QWQi,KWKi,VWVi) and Wo ​ is the output projection matrix.

3.2.4 Feedforward neural network

After the attention process, the model uses feedforward neural networks to perform further nonlinear transformations. The FFN is made up of two linear layers separated by a non-linear activation function (often a ReLU). The mathematical expression for the FFN operation is as follows:

FFN(x)=ReLU(xW1+b1)W2+b2(7)

The weight matrices and bias terms are represented by  W1, b1, W2 and b2.

In the context of ASD detection, the Transformer encoder examines neuroimaging data, such as MRI scans, to identify ASD-related characteristics and patterns. The self-attention mechanism allows the model to focus on specific areas of interest, whilst the multi-head mechanism improves the model’s capacity to collect varied and complicated correlations in brain data. The ensuing feed forward neural networks modify the representations for downstream tasks, such as distinguishing between ASD and control participants.

4 Proposed model of ASD multi-view united transformer block

In this section, we introduce the ASD Multi-View United Transformer Block (ASD-MVUTB), a new model for detecting Autism Spectrum Disorder (ASD). Our suggested design uses transformer blocks to combine multi-view information, taking use of varied views from neuroimaging data. The model effectively captures subtle patterns from several perspectives, encouraging complete feature extraction for accurate ASD detection. Experimental validation shows that the proposed ASD-MVUTB outperforms existing techniques.

The workflow for detecting Autism Spectrum Disorder (ASD) in Figure 1, is structured around N distinct views (V1, V2, and VN), each comprising both positive (+V) and negative (-V) perspectives. In the V1-Positive View (+V), the process initiates with a Graph Attention Network (GAT-1), wherein neuroimaging data undergoes attention-based feature extraction. The output of GAT-1 is then subject to a Residual Addition operation, combining the GAT-1 output and the original positive view data. Subsequently, another Graph Attention Network (GAT-2) is applied, repeating the attention-based feature extraction process. The results of GAT-2 for both +V and -V are concatenated, and the concatenated output undergoes two linear layers (linear (128,16) and linear (16,2)) to determine the binary classification output, indicating whether the patient exhibits ASD (39, 40).

Figure 1. Workflow of detecting ASD.

Mathematically, the process can be expressed as follows:

1. GAT-1 Process:

GAT−1(+V)=GAT−1 Output(8)

2. Residual Addition:

GAT−1 Output++V=Residual Addition(9)

3. GAT-2 Process:

GAT−2(Residual Addition Output)=GAT−2 Output(10)

4. GAT-2 Layer Concatenation:

Concatenate(GAT−2 Output (+V), GAT−2 Output (−V))(11)

5. Linear Layer Application:

Linear (128, 16)→Lineary (16, 2)=View1 Output(12)

Next, after the GAT-2 process for each view, a Transformer Encoder (TE) is applied. The TE operation is conducted separately for both the +V and -V outcomes of each view. This multi-view approach enhances the model’s ability to capture diverse features and patterns related to ASD within the neuroimaging data.

The Transformer Encoder for all View (V’s) is expressed as:

TE(GAT−2 Output(+V))=+V TE Output(13)TE(GAT−2 Output(−V))=−V TE Output(14)

This entire process is replicated for View2 (V2), View3 (V3) and upto ViewN(VN). In this work three views are used to construct the ASD model. The utilization of Graph Attention Networks and Transformer Encoders facilitates capturing intricate patterns in neuroimaging data, and the linear layers aid in synthesizing and classifying the information for ASD detection. The concatenated results from the GAT-2 process ensure a comprehensive representation of both positive and negative perspectives in each view, contributing to the overall diagnostic outcome.

The proposed ASD-MVUTB model architecture, as illustrated in Figure 2, involves distinct processing of positive and negative TE outputs from Views (+Vs and -Vs) separately. Specifically, the TE outputs +ViTEout (X1) and +Vi+1TEout (X2) are processed independently in the first MVUTB layer. The resulting outputs, denoted as X1 and X2, are then combined with +Vi+2TEout in the second MVUTB layer. These intermediate results are flattened to create feature vectors. A parallel process is applied to the negative TE outputs, resulting in three sets of flattened feature vectors. Subsequently, the flattened results from both +Vs and -Vs are concatenated, and the concatenated feature vector undergoes three linear transformations: Linear(384, 32), Linear(32, 16), and Linear(16, 2).

Figure 2. ASD-MVUTB proposed model architecture.

The ASD-MVUTB working process is represented as follows:

1. MVUTB layers:

• Positive TE Outputs Processing.

X1=MVUTB(+ViTEOUT, +Vi+1TEOUT)(15)X2=MVUTB(X1, +Vi+2TEOUT)(16)

• Negative TE Outputs Processing.

Y1=MVUTB(−ViTEOUT, −Vi+1TEOUT) (17)Y2=MVUTB(Y1, −Vi+2TEOUT)(18)

2. +Vs and –Vs Flatteninng.

3. Concatenation:

FConcat=Concat(FX, FY)(21)

4. Linear Transformations:

Z1=Linear (FConcat, 384,32)(22)Z2=Linear (Z1,32,16)(23)Z3=Linear (Z2, 16, 2)(24)

The motivation behind this architecture lies in the enhanced feature extraction from both positive and negate+ve TE outputs independently, and the subsequent concatenation allows the model to capture and leverage information from both perspectives. The linear layers at the end serve to further distill and map the concatenated features into a final prediction space for ASD classification. This layered approach enhances the model’s ability to discern intricate patterns and relationships within the TE outputs, contributing to its effectiveness in ASD recognition.

In the proposed framework depicted in Figure 3, the Multi-View Transformer Block operates through a meticulous process involving two views of features, denoted as X1 and X2. These views undergo Layer Normalization simultaneously. Subsequently, a linear mapping is applied to the query (q) and key (k) from X1, and the value (v) from X2, which are then fed into the Multi-Head Attention (MHA) mechanism. The output of MHA, along with the original features from X1, undergoes an element-wise addition (A1) to yield a preliminary result (R1). Following this, another Layer Normalization is applied, succeeded by a Multi-Layer Perceptron (MLP). The output of the MLP, combined with the result from the previous addition (R1), undergoes a second addition (A2), culminating in the generation of the final outcome. This iterative and additive process allows for the effective integration of information from multiple views.