Autism spectrum disorder identification with multi-site functional magnetic resonance imaging

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 2, June 2024, pp. 2143~2154
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i2.pp2143-2154  2143
Journal homepage: http://ijai.iaescore.com
Autism spectrum disorder identification with multi-site
functional magnetic resonance imaging
Shabeena Lylath, Laxmi B. Rananavare
Department of Computer Science and Engineering, Reva University, Bangalore, India
Article Info ABSTRACT
Article history:
Received Aug 3, 2023
Revised Oct 8, 2023
Accepted Jan 6, 2024
Autism spectrum disorder (ASD) is a neurodevelopmental condition
characterized by enduring difficulties in social interaction and
communication. People analyzed with ASD may display repetitive behaviors
and limited interests. Autism is classified as a spectrum disorder, implying
that the symptom intensity might range from mild to severe depending on
the individual. To detect ASD in this paper an attribute feature graph
approach is designed by using the stastical dependencies features that
necessarily accomplish the diagnosis of ASD. In the first phase the features
extracted are designed based on the functional magnetic resonance imaging
(fMRI) data, in the next-step the attribute feature graph layer learns the
features of the node information of various nodes by ASD classification.
Further, in the third step, it is employed to independently extract
distinguishing features from the functional connectivity matrices of the brain
that are derived from fMRI. The custom convolutional neural network
(CNN) used in this study is trained on a comprehensive dataset comprising
individuals diagnosed with ASD and typically developing individuals. In the
fourth stage, a prototype learning is developed to augment the classification
performance of the custom-CNN. The experimental analysis further carried
out states that the proposed model works efficiently in comparison with the
existing system.
Keywords:
Attribute feature graph
Autism spectrum disorder
Convolutional neural networks
Functional magnetic resonance
imaging data
Neurodevelopmental condition
This is an open access article under the CC BY-SA license.
Corresponding Author:
Shabeena Lylath
Department of Computer Science and Engineering, Reva University
Bangalore, India
Email: shabeenal_12@rediffmail.com
1. INTRODUCTION
Autism spectrum disorder (ASD) is a neurodevelopmental disease that has detrimental effects on
early brain development and an individual’s capacity for social interaction [1]. ASD is characterized by the
presence of restrictive and repetitive behavioral patterns. The term “spectrum” encompasses a diverse range
of symptoms that vary in their level of severity [2]. Although there is currently no established treatment for
ASD, early intervention and appropriate medical care have the potential to significantly improve a child’s
development by enhancing their behavioral and communicative skills [3]. Identifying and diagnosing ASD
through standard behavioral investigations pose significant challenges and complexities. Autism may be
diagnosed beyond the typical age of two, depending on it is severity [4]. There exist multiple accessible
treatment strategies that can be employed to detect ASD at the earliest possible stage. Diagnostic procedures
are typically not extensively utilized until a child is deemed to have a high risk of developing autism.
The diagnosis of ASD is typically made in children aged 48 to 60 months, indicating a delay of two
years [5]. Most children are typically not diagnosed with ASD until they reach the age of 48 to 60 months.
This is the case even though there are observable signs of ASD as early as 12 months, and there is a

 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 2, June 2024: 2143-2154
2144
possibility of diagnosis by 18 months [6]. The diagnostic delay of 12 to 14 months has an impact on early
intervention programs, which have the potential to increase children’s IQs by 10 to 15 points if initiated
before the age of three [7]. In order to enhance the quality of life for children diagnosed with ASD, it is
imperative to prioritize early diagnosis and treatment for ASD.
Magnetic resonance imaging (MRI) and positron emission tomography (PET) are commonly
utilized neuroimaging techniques for investigating the neurodevelopmental characteristics associated with
ASD [8]. In recent decades, there has been a notable increase in research dedicated to investigating structural
and functional brain abnormalities that may serve as indicators of ASD. The neurodevelopmental
characteristics of ASD have been extensively studied in MRI investigations. It is important to note that while
these investigations have revealed numerous consequences, it is crucial to recognize that the findings may not
universally apply to all individuals with ASD. Functional magnetic resonance imaging (fMRI) investigations
primarily focus on examining the local and global connectivity patterns within the brain. In contrast,
structural MRI studies commonly employ volumetric and morphometric analysis techniques to investigate
any deviations or abnormalities in the structure of the brain.
The absence of a standardized test or tool [9] for the diagnosis of autism presents a significant
challenge in achieving an accurate diagnosis. Physicians conduct behavioral observations of the children or
inquire with the parents regarding the children’s medical history. One of the two diagnostic methods utilized
for the identification of autism is the autism diagnostic observation schedule (ADOS). This particular method
places it is emphasis on the assessment of social interaction. The evaluator closely observes the individual
under assessment and subsequently determines a grade based on their observations. During the assessment
process for autism, the subject is subjected to observation, and the examiner assigns scores based on their
observations. The autism diagnostic interview-revised (ADI-R) is a commonly utilized structured interview
that involves gathering information from parents regarding the developmental history of the individual [10].
The diagnostic tools used in this context have an observational nature, which increases the likelihood of
generating false-positive results. This is especially true for individuals with a range of mental health
disorders. The examples that deserve special attention include cases involving psychotic patients [11],
childhood-onset schizophrenia [12], attention-deficit hyperactivity disorder (ADHD) [12], and various other
disorders. The pursuit of precise brain biomarkers and the automation of the identification process for ASD
are essential goals in diagnostic procedures. These objectives aim to overcome the limitations associated with
clinical tests.
Deep learning algorithms have demonstrated significant success in the field of image processing. As
a result, there is a considerable amount of interest in leveraging these algorithms for the analysis of fMRI
data. This study examines the challenges and potential applications of deep learning techniques in the
analysis of fMRI data for the detection and classification of neurological disorders. To identify ASDs,
researchers have examined fMRI data using various deep learning techniques, including convolutional neural
networks (CNNs). CNNs are particularly effective as they utilize local spatial patterns and hierarchical
representations of brain activity [13].
Research studies have demonstrated that CNN-based methods exhibit the capability to accurately
classify ASDs by effectively identifying distinctive characteristics specific to ASDs. Recurrent neural
networks (RNNs) have been employed for the estimation of time-varying brain activity associated with ASD.
This is achieved by leveraging the RNN’s ability to capture the inherent temporal dynamics present in fMRI
data. RNNs have the capability to identify patterns that differentiate individuals with ASD from those with
typical brain development. This is achieved by analyzing sequential information present in fMRI data.
Autoencoders have been employed for reducing the complexity of fMRI data, facilitating the identification of
significant features for ASD classification. The reduction of noise and excessive dimensionality observed in
fMRI data facilitates the identification of patterns associated with ASD. Despite the promising outcomes,
there are several challenges that need to be addressed before deep learning methods can be effectively
employed for ASD identification using fMRI data [14], [15]. The motivation and contribution for this
research is mentioned below.
− ASD has emerged as a significant public health concern due to its escalating prevalence. Efforts should be
made to enhance the welfare of individuals affected by this condition through timely identification and
effective treatment. One potential approach for accurately identifying ASD and advancing our
understanding of its neurological underpinnings involves the application of deep learning techniques in
the analysis of fMRI data [16]. The primary objective of this study is to facilitate the advancement of
deep learning applications in neuroimaging by tackling challenges related to data quality, pre-processing,
model interpretability, and restricted data accessibility. The objective of the study is to facilitate
collaboration and facilitate the sharing of data within the industry. Ultimately, the utilization of deep
learning techniques for fMRI data analysis in the context of ASD diagnosis holds promise for enhancing
clinical outcomes and enhancing the quality of life for individuals affected by this disorder.

Int J Artif Intell ISSN: 2252-8938 
Autism spectrum disorder identification with multi-site functional magnetic resonance … (Shabeena Lylath)
2145
− In this paper an attribute feature graph approach is designed by using the stastical dependencies features
that necessarily accomplish the diagnosis of ASD. In the first phase the features extracted are designed
based on the fMRI data with each section that is developed as a graph, in the next-step the attribute
feature graph layer learns the features of the node information of various nodes by ASD classification.
− The main objective of this study is to use fMRI data to create a classification model that can distinguish
between individuals diagnosed with ASD.
− The proposed model utilizes a transfer learning technique to improve model design and enhance
classification performance. The researchers conducted systematic testing on the autism brain imaging data
exchange (ABIDE) dataset, demonstrating that the proposed model outperforms existing state-of-the-art
techniques efficiently.
Advancements in deep learning have enabled the integration of various data types, including
multimodal data [17], [18]. Several academic studies have shown the effectiveness of deep learning in
addressing medical challenges, including the identification of Alzheimer’s disease using MRI and PET data
[19], [20]. Deep learning has demonstrated success in understanding complex patterns like functional
connectivity, making it a valuable diagnostic tool [21]. Utilizing multiple voxels as inputs allows exploring
complex associations among variables at a higher level, leading to discrimination between affected and
healthy cohorts and proposing optimal diagnostic strategies [22]. According to Nie et al. [22], a method
called transfer subspace learning via low rank and sparse representation (TSL_LRSR) was proposed for
unsupervised domain transfer, demonstrating promising outcomes on visual domain adaptation tasks.
Variants of low rank representation (LRR), such as multi-site adaption framework via low-rank
representation (maLRR) [23] and multi-source domain adaptation (MSDA) and multi-view sparse
representation (MVSR), have been employed for identifying ASD across multiple sites, enhancing the
precision of ASD classification by mitigating data heterogeneity.
According to Zhang et al. [13], axial slices of T1-weighted (T1w) sMRI were used for detecting
anomalies through a two-step approach, achieving high area under curve (AUC) scores with generative
adversarial network (GAN) architecture. According to Tang et al. [17], self-attention modules in GAN
architectures were investigated for detecting alzheimer's disease and brain metastases, leading to significant
AUC values. The abnormal-to-normal translation generative adversarial network (ANT-GAN), a cycle
generative adversarial network (CycleGAN)-based approach, was devised to generate medical images to
distinguish healthy and abnormal scans. Our study presents a novel approach [20] for multi-site adaptation
using fMRI and low-rank representation (LRR) decomposition, specifically the maLRR framework. The
primary goal is to identify a shared LRR for data from multiple sites, mitigating distribution disparities
between them. The non-oscillatory brain connectivity technique [23] has been proposed to distinguish
between ASD subtypes based on resting-state functional magnetic resonance imaging (rs-fMRI) signal.
The research work in this paper is organized in four section, the first section deals with the
background of ASD its challenges and the related work is written involving the existing techniques used
involving various deep learning techniques. In the second section, the proposed methodology is designed for
ASD diagnosis. In the third section the results are evaluated on the ABIDE dataset which further evaluated
various performance metrics considered.
2. METHOD
The proposed model involves constructing brain statistical dependencies for spectrum disorder
(SD)-based classification, which includes network deployment, feature learning (feature extraction and
classification). Currently, the statistical dependencies are built from pre-processed MRI data using correlation
analyses. The framework deals with high-dimensional features and a limited sample size of MRI data. To
address the issue of inappropriate features, the rank correlation coefficient is utilized to develop an input
feature map, emphasizing the most relevant features. The feature extractor employs a primitive CNN model,
where various prototypes are automatically learned to represent different categories. A generic prototype
loss, based on cross-entropy, is designed to optimize the CNN feature extraction and prototypes.
Classification is achieved by matching and estimating Euclidean distance. Additionally, domain adaptation is
introduced to enhance model training. Figure 1 shows the proposed framework.
The statistical dependencies analysis method is evaluated to analyse pairwise correlations amongst
the median value of the MRI data utilized in every brain region. The correlation is used to evaluate the length
of MRI time-series. Each subject here is defined as 𝑔𝑥(𝑢), 𝑔𝑦(𝑢) ∈ 𝛽𝑍
. The average of the MRI signals for
the brain section 𝑥 and 𝑦 at time period 𝑢(𝑢 = 1,2, … . , 𝑈). 𝑍 and 𝑈 depicts the number regions of interest
(ROIs). with total number of points respectively. Then 𝑆𝐷𝑥𝑦 is defined as given in (1). Here 𝑔𝑥 and 𝑔𝑦 depict
the mean of the MRI signals in the brain section 𝑥 and 𝑦. An SD network is constructed using correlation-
based methods by evaluating the Pearson correlation between the time-series of each pair of ROIs.

 ISSN: 2252-8938
2146
𝑆𝐷𝑥𝑦 =
∑ (𝑔𝑥(𝑢)−𝑔𝑥)(𝑔𝑦(𝑢)−𝑔𝑦)
𝑈
𝑢=1
√∑ (
𝑈
𝑢=1 𝑔𝑥(𝑢)−𝑔𝑥)2∗√∑ (
𝑈
𝑢=1 𝑔𝑦(𝑢)−𝑔𝑦)2
(1)
Figure 1. Proposed framework
2.1. Correlation estimation
The proposed framework utilizes rank correlation to generate input feature maps from
high-dimensional MRI data samples. The feature ranking technique is critical for identifying crucial and
pertinent MRI data and minimizing irrelevant data that slows down processing. To enhance 𝑔𝑥𝑦 which
denotes the 𝑥 − 𝑡ℎ statistical dependencies for the 𝑦 − 𝑡ℎ sample and the ℎ𝑦 shows the sample’s class label.
The observation pairs shown here consists of the datasets {𝑔𝑥𝑦, ℎ𝑦} and {𝑔𝑥𝑖, ℎ𝑖} in in the distribution of (2).
µ is a signum function, alternately {𝑔𝑥𝑦, ℎ𝑦} is in distribution of {𝑔𝑥𝑖, ℎ𝑖} where, shown as in (3). The
correlation coefficient (𝝳) for the 𝑥 − 𝑡ℎ 𝑆𝐷 is defined as in given (4). In the below equation 𝑘𝑎 and 𝑘𝑏
depicts the amount of distributions and pairs irrespective of the (2) and (3), 𝑧 and 𝑙 depicts how many
samples were collected from the patient and control groups. The 𝑆𝐷 is ranked in accordance with (δ), the
input-feature map is projected by the proposed framework developed by the 𝑆𝐷.
µ(𝑔𝑥𝑦 − 𝑔𝑥𝑖} = µ(ℎ𝑦 − ℎ𝑖) (2)
µ(𝑔𝑥𝑦 − 𝑔𝑥𝑖} = −µ(ℎ𝑦 − ℎ𝑖) (3)
δ𝑥 =
𝑘𝑎−𝑘𝑏
𝑧∗𝑙
(4)
2.2. Proposed framework
The neural network automatically learns multi-level features from the input training data using
filters and updates its parameters during the training process. In this research, a novel artificial neural
network (ANN) integrated with the proposed framework for classifying the brain’s functional network. The
ANN here is employed through the feature extraction mechanism depicted as ℎ(𝑔; 𝜑), here 𝑔 denotes the
input samples whereas 𝜑 the parameters of ANN. In this model, the ANN consists of an input layer, three
convolution layers, a max-pooling layer (2×2), a fully connected layer, and an output layer. The max-pooling
layer is utilized for down-sampling and selecting values within local neighborhoods. The activation function
𝑅 is employed.
In the context of prototype learning, the output of the feature extraction step is utilized from both
nodes. The prototypes for each class were created using the basic characteristics of the ANN in a fixed
manner. This strategy introduces a novel technique in which prototypes are automatically learned for each
category. The creation of each prototype begins by utilizing the initial prototype specific to its corresponding
class. This initial prototype is developed based on the feature space derived from a pre-training set, which
captures the characteristics present in each sample category. The distance between each sample and the
current prototype is calculated by considering the adjacent prototypes at each sample. The attributes that have
been identified through multiple prototypes associated with each category are subsequently utilized in the
development of the subsequent prototype for that particular category. Here 𝑈 = {𝑢𝑥𝑦|𝑥 = 1,2, … … , 𝐺;
𝑦 = 1,2, … . . , 𝐼} depicts the prototypes where 𝑥{1,2,……,G} represents the index for the categories and
𝑥  {1,2, … … , 𝐼} shows the prototype index for each category that is displayed. The model training involves,
the feature extractor ℎ(𝑔; 𝜑), the prototypes as {𝑢𝑥𝑦} are trained combined. For model training, a prototype
loss function is used, which is based on distance-based cross-entropy. This loss function helps optimize the
model by taking into account the differences between the prototypes and the training data. The
distance-entropy generated during model training serves as the metric for measuring the similarity between
samples and is utilized as the foundation for the loss function (𝑔, ℎ) and the prototype 𝑢𝑥𝑦.

2147
𝑝(ℎ(𝑔; 𝜑), 𝑢𝑥𝑦)=|| ℎ(𝑔; 𝜑) − 𝑢𝑥𝑦|| (5)
The likelihood or probability of the sample (𝑔, ℎ) the association or belongingness of the sample to
the prototype. 𝑢𝑥𝑦 can be estimated by 𝐶(𝑔 ∈ 𝑢𝑥𝑦|𝑔) ∝ −𝑝(ℎ(𝑔; 𝜑), 𝑢𝑥𝑦). To guarantee the non-negativity
and the probability total as 𝐶(𝑔 ∈ 𝑢𝑥𝑦|𝑔) is defined aa in (6). 𝜇 is a hyper parameter, through the probability
𝐶(𝑔 ∈ 𝑢𝑥𝑦|𝑔) then 𝐶(𝑔|ℎ) is defined as given in (7). The cross-entropy loss is defined as the negative log in
accordance with the (8).
𝐶(𝑔 ∈ 𝑢𝑥𝑦|𝑔) =
exp (−𝜇𝑝(ℎ(𝑔;𝜑),𝑢𝑥𝑦)
∑𝐵
𝑜=1 ∑ exp (−𝜇𝑝(ℎ(𝑔;𝜑),𝑢𝑥𝑦))
𝐼
𝑚=1
(6)
𝐶(𝑔|ℎ) = ∑ 𝐶(𝑔 ∈ 𝑢𝑔𝑦|ℎ)
𝐼
𝑦=1 (7)
𝑙𝑜𝑠𝑠((𝑔, ℎ); 𝜑, 𝑈) = −𝑙𝑜𝑔𝐶(𝑔|ℎ) (8)
To reduce the distance-based entropy the loss is decreased and the separation between the samples
and the category prototype. The feature learnt is closer to the class in improvisation of intra-class concise,
causing the feature representation to stand out. The prototype loss further lowers overfitting, improving
generalizability in the process. The prototype loss is defined as in (9). Here 𝑢𝑔𝑦 denotes the closest prototype
for ℎ(𝑔; 𝜑) for the corresponding section 𝑔. The distance based-loss with the prototype to determine the total
loss is mentioned in (10).
𝑐𝑙𝑜𝑠𝑠((𝑔, ℎ); 𝜑, 𝑈) = ||ℎ(𝑔; 𝜑) − 𝑢𝑔𝑦||2
(9)
𝑇𝑜𝑡𝑎𝑙(((𝑔, ℎ); 𝜑, 𝑈)
= 𝑙𝑜𝑠𝑠((𝑔, ℎ); 𝜑, 𝑈)+𝜗 𝑐𝑙𝑜𝑠𝑠((𝑔, ℎ); 𝜑, 𝑈)
−𝑙𝑜𝑔 ∑
exp (−𝜇𝑝(ℎ(𝑔;𝜑),𝑢𝑥𝑦)
∑𝐵
𝑜=1 ∑ exp (−𝜇𝑝(ℎ(𝑔;𝜑),𝑢𝑜𝑚))
𝐼
𝑚=1
𝐼
𝑦=1 + 𝜗||ℎ(𝑔; 𝜑) − 𝑢𝑔𝑦||2
(10)
𝜗 is the hyper-parameter for the 𝜑 and the prototype are trained by minimising the loss function in (10). This
is further computed in accordance with the error propagation algorithm and derive in regard with the updated
parameters is determined by the chain rule. The optimization algorithm is employed for training the model
for updating the framework. The number of iterations allowed throughout the training phase is set at 800.,
with initial learning fixed to 0.0001 for the hyper parameter 𝜗. Algorithm 1 is mentioned as follows.
Algorithm 1. Optimization algorithm
Step 1 𝐷𝑎𝑡𝑎𝑠𝑒𝑡={(𝑔1,ℎ1), (𝑔2, ℎ2), … … . , (𝑔𝑘, ℎ𝑘)), loss denoted by 𝑜
Step 2 Initialization for the feature extraction ℎ(𝑔; 𝜑) and the prototypes {𝑢𝑥𝑦}
Step 3 While stopping criteria not meet do
Step 4 Sampling batch 𝑢 from 𝐷𝑎𝑡𝑎𝑠𝑒𝑡
Step 5 Batch u sample into the proposed forward propagation
Step 6 The gradients evaluated are
𝑑𝑜
𝑑𝜑
and
𝑑𝑜
𝑢𝑥𝑦
by back propagation and chain rule
Step 7 Update the value 𝜑 and {𝑢𝑥𝑦} the Adam optimization algorithm utilizes gradients, as per its formulation.
𝑑𝑜
𝑑𝜑
and
𝑑𝑜
𝑢𝑥𝑦
Step 8 end while
Step 9 return updated model parameters 𝜑 and {𝑢𝑥𝑦}
In the classification stage, the sample is classified based on prototype matching, and it is identified
as belonging to the adjacent prototype through the calculation of Euclidean distance in the feature space. This
process is defined as given in (11). Here 𝑣𝑥(𝑔) is the discriminant function in the section 𝑥, as mentioned in
(12). The input pattern, represented as 𝑔 in abstract form, is compared to a number of prototypes, and the
input sample is categorized according to the location of the neighboring prototype.
𝑔𝜖 sec 𝑎𝑟𝑔𝑚𝑖𝑛 𝑣𝑥(𝑔)
𝑥=1
𝐵
(11)
𝑣𝑥(𝑔) = 𝑚𝑖𝑛 ||ℎ(𝑔; 𝜑)||2
𝑦=1
𝐼
(12)
2.3. Developing statistical dependencies features
Statistical dependencies features are constructed by assessing the similarity on the pre-processed
data. The implementation procedure includes segmenting the brain into multiple divisions, and in the second
step, the correlation coefficient is employed to estimate the correlation 𝐴𝑥𝑦 amidst the brain segments 𝐴𝑥 and
𝐴𝑦. Finally, a correlation co-efficient matrix is determined for Z*Z matrix built for each subject. Here Z

 ISSN: 2252-8938
2148
denotes the ROI. The statistical dependencies matrix, is expressed as given in (13). Each row in the vector 𝑗𝑎
for the statistical dependencies matrix is selected for the SD features shown as in (14).
𝐺 =
𝐴11 𝐴12 𝐴1𝑍
𝐴21 𝐴22 𝐴2𝑍
𝐴𝑍1 𝐴𝑍2 𝐴𝑍𝑍
(13)
𝑗𝑚 = 𝐴𝑚1,𝐴𝑚2, … . , 𝐴𝑚𝑧 (14)
2.4. Attribute feature graph with self-attention mechanism layers
The integration of the attribute feature graph model with statistical dependencies features involves
two segments. In the first section, an ROI is used as a node, and the statistical dependencies features
extracted from the MRI data are combined into the graph. The connections linking the nodes in the graph,
which reflect the characteristics of each subject, create the graph. The graph representation for each subject is
denoted as 𝐺 = 𝑗, 𝐹 here 𝑏 is the node feature set and 𝐹 The edges connect the two nodes in the graph. The
second part deals with node features and the process of detecting ASD. Figure 2 shows the workflow of the
proposed architecture.
− Designing an attribute feature graph: In this representation, each ROI is denoted as a node, The absolute
number denotes the weight assigned to each edge as 𝐴𝑥𝑦: |𝐴𝑥𝑦|, the edge set is expressed as:
𝐹 = |𝐴11|, | 𝐴12|, … , |𝐴𝑥𝑦|, 𝑥, 𝑦 ∈ 1,110. The designed feature-set for each row of statistical dependency
is used as an attribute feature graph to generate = 𝑗1, 𝑗2, … , 𝑗𝑧 … . , 𝑗𝑥 ∈ 𝛽𝑛𝑎
, 𝑛𝑎 . The node dimension
depicts the attributes of each node. The attribute feature graph is built for each subject as 𝐺 = 𝑗, 𝐹.
− Self-attention-based graph layers: this is developed for the purpose of learning node representation. Here
the 𝑍 value of attribute feature graph is fixed to 2. These node’s input characteristics is 𝑗 =
𝑗1, 𝑗2, … , 𝑗𝑧 … . , 𝑗𝑥 ∈ 𝛽𝑛𝑎
and the output of the for the node characteristics, node-set 𝑗′
=
𝑗1
′
, 𝑗2,
′
… , 𝑗𝑧
′
… . , 𝑗𝑥
′
∈ 𝛽𝑛𝑎′
. A weight matrix 𝑊𝑀𝜖𝛽𝑛𝑎′∗𝑛𝑎
trained amongst other nodes. The nodes are
shown, and a self-attention process aggregates the nearby nodes for each node. The purpose of the
attention coefficient is to:
𝑘𝑥𝑦 = 𝑣𝑊𝑀𝑗𝑥, 𝑊𝑀𝑗𝑦 (15)
Here 𝑣: 𝛽𝑛𝑎′
*𝛽𝑛𝑎′
→β depicts the attention model which includes self-attention along with
neighbourhood attention as 𝑘𝑥𝑦 that depicts the significance of node 𝑦 to 𝑥. The softmax function denoted by
is introduced that regularizes the neighbouring nodes denoted by 𝑦 ∈ 𝑍𝑥 of, as shown in (16). The attention
mechanism 𝑣 is a neural network, 𝑣 ∈ 𝛽2𝑛𝑎′
. The sigmoid function in the output layer of the neural network
is coupled to the weight matrix. The attention-based cross-correlation coefficient is obtained via (17).
Here 𝑡𝑟𝑎𝑛𝑠 depicts matrix transposition. The layers of the neural network, including the output layer’s
sigmoid function, are linked to the weight matrix. The attention-based cross-correlation coefficient is
obtained shown as in (6). 𝛿 denotes the activation function and 𝑦 is traversed in 𝑦 ∈ 𝑍𝑥 denotes all nodes
parallel to 𝑥 as given in (18).
𝑛𝑥𝑦 =ꭈ 𝑘𝑥𝑦 =
exp 𝑘𝑥𝑦
∑ exp 𝑘𝑥𝑦
𝑖∈ 𝑍𝑥
(16)
𝑛𝑥𝑦 = ꭈ𝑠𝑖𝑔𝑘𝑥𝑦
exp 𝑠𝑖𝑔𝑣𝑡𝑟𝑎𝑛𝑠[𝑊𝑀𝑗𝑥||𝑊𝑀𝑗𝑦]
∑ exp 𝑠𝑖𝑔𝑣𝑡𝑟𝑎𝑛𝑠[𝑊𝑀𝑗𝑥||𝑊𝑀𝑗𝑖]
𝑖∈ 𝑍𝑥
(17)
𝑗𝑥
′
= 𝛿[∑ 𝑛𝑥𝑦
𝑦∈ 𝑍𝑥
𝑊𝑀𝑗𝑦] (18)
Figure 2. Workflow of the proposed architecture

2149
2.5. Attention mechanism
An attention mechanism is built by the transformer attention mechanism. The features are learnt by
this and they describe signals through various sides. The independent attention mechanisms are described in 𝐼
in (7), the nodes in feature representation depicts each attention to develop the final node representation,
represented as given in (19).
𝑗𝑥
′
=
𝑖=1
||
𝐼
𝛿[∑ 𝑛𝑥𝑦
𝐼
𝑦∈ 𝑍𝑥
𝑊𝑀𝐼
𝑗𝑦] (19)
Here I denote the attention heads, 𝐼 = 3. 𝑛𝑥𝑦
𝐼
is the attention co-efficient estimated through the 𝐼 − 𝑡ℎ
attention head with the size of
𝑛𝑎′
𝐼
∗ 𝑛𝑎. We immediate layer for node feature representing by each attention
connection via the output dimension for 𝐼𝑛𝑎′
. The attribute feature graph performs multiple attention
mechanisms, 𝐼-th in the final layer, the attention-based cross-correlation coefficient is utilized instead of the
operation to generate the final output, preserving the initial input dimension. Along with for allowing the
application of the final non-linear function. The node feature representation in the final layer is depicted as
given in (20). The attention mechanism updates the node𝑗1 representation in the neighborhood based on the
calculated weights from the target node.
𝑗𝑥
′
= 𝛿[
1
𝐼
∑ ∑ 𝑛𝑥𝑦
𝐼
𝑦∈ 𝑍𝑥
𝑊𝑀𝐼
𝑗𝑦]
𝐼
𝑖=1 (20)
3. PERFORMANCE EVALUATION
To assess the effectiveness of the proposed method, the study conducts inter-site cross-validation
using data from 17 sites obtained from the ABIDE database. During the validation process, one site is
designated as the testing target domain, while the remaining 16 sites act as multiple training source domains.
This procedure is repeated multiple times, allowing each site to serve as the target domain to ensure the
model’s stability and generalizability in handling heterogeneous multi-site data.
3.1. Dataset details
The ABIDE collection comprises data from individuals diagnosed with ASD as well as individuals
who exhibit typical development. The availability of this dataset enables researchers to conduct studies on the
neurological foundations of ASD, owing to its comprehensive and diverse assortment of imaging and clinical
data. The dataset comprises structural MRI, resting-state fMRI, and diffusion-weighted magnetic resonance
imaging (dMRI) scans for each participant, along with accompanying clinical and demographic information.
A comprehensive dataset comprising data from multiple imaging modalities and scanners has been generated
through collaborative efforts from various research institutes worldwide. The ABIDE dataset is a valuable
resource for researchers interested in exploring the connections between brain connectivity, structural
abnormalities, and biomarkers associated with ASD. The utilization of this tool has been extensive in the
enhancement and testing of machine learning and data-driven approaches for the diagnosis and description of
autism. In this study, the ABIDE I dataset serves as the primary focus of analysis, encompassing information
from a total of 1,102 participants.
3.2. Comparison methods considered on the ABIDE dataset
The suggested method’s effectiveness is assessed by comparing the findings with those from
different other disciplines, and the effectiveness is further demonstrated by experimental analysis. The
ABIDE collection comprises data from individuals diagnosed with ASD as well as individuals who exhibit
typical development. A comprehensive dataset comprising data from multiple imaging modalities and
scanners has been generated through collaborative efforts from various research institutes worldwide.
− Correlation alignment (CORAL) [24]: the goal of the approach is to address domain shift by aligning the
second-order statistics of the probability distribution in both the source and target domains.
− Joint distribution adaptation (JDA) [25]: concurrently adjusting the marginal and conditional distributions
during dimensionality reduction narrows the distributional gap between domains, thus facilitating the
learning of a new data feature representation.
− Transfer component analysis (TCA) [26]: the approach involves generating a cross-domain feature
representation using feature dimensionality reduction and the greatest mean discrepancy to measure the
distance between domains.
− Manifold embedded distribution alignment (MEDA) [27]: the employed approach utilizes manifold
feature learning and dynamic distribution alignment techniques to assess the relative importance of the

 ISSN: 2252-8938
2150
marginal and conditional distributions. This enables the establishment of a domain-invariant classifier
capable of handling variations across domains.
− Transfer subspace learning via low rank and sparse representation (TSL_LRSR) [28]: the proposed
methodology incorporates the use of LRR to project data from two separate domains onto a shared
subspace. This enables the extraction of domain-invariant characteristic representations, effectively
addressing distributional disparities between the two domains.
− Centroid matching and local manifold self-learning (CMMS) [29]: this method aims to minimize the
disparity in the conditional distribution between two domains by utilizing a local manifold self-learning
strategy to effectively capture the inherent local manifold structure of the target domain samples. This
involves adaptively identifying centroids in both domains, enabling the direct assignment of labels to the
target domain data. In this study, potential optimization parameters are investigated within a specific set
of values: 0.001, 0.01, 0.1, 1, 10, and 100. Notably, default parameters from the references are excluded
from this search. The following section outlines the advanced multi-site techniques used in the
comparative analyses for detecting ASD.
− maLRR [30]: the domain invariant representation of features is obtained using LRR. This involves
extracting the shared component from specific projection matrices related to different source domains,
which are then utilized as the reference domain’s projection matrix.
− Denoising autoencoder [31]: the method involves reducing the dimensionality of the data by learning a
domain-invariant feature representation unsupervisedly using two stacked denoising autoencoders.
− Deep belief network (DBN) [32]: this study employs a depth-first search methodology with a constrained
route to investigate the topological information concealed within a graph. The selection of significant
functional connection characteristics is accomplished through the utilization of a graph-based feature
selection approach. Subsequently, an ASD identification process is carried out through the utilization of a
three-tiered DBN that incorporates automated hyper-parameter modification.
− Autism spectrum disorder framework for diagnostic network (ASD-DiagNet) [33]: the utilization of
autoencoder and single-layer perceptrons in combination with joint learning is employed to augment the
precision The utilization of feature extraction enhances the model’s parameters.
− Federated training and adversarial alignment approach (Fed_Align) [34]: This approach showcases the
technique of utilizing the adversarial domain adaptation approach to depict domain-invariant
characteristics.
3.3. Results
The results are evaluated for the proposed method for accuracy, sensitivity, specificity and recall
and depicted in the form of graph for various existing state-of-art methods CORAL [24], JDA [25], TCA
[26], MEDA [27], TSL_LRSR [28], CMMS [29], maLRR [30], denoising autoencoder [31], DBN [32],
ASD-DiagNet [33], Fed_Align [34]. Table 1 shows the accuracy comparison for various state-of-art methods.
From the accuracy evaluation, we can say that proposed system has better accuracy. In Figure 3, the accuracy
comparison is carried out with various existing state-of-art methods. Where maLRR [30] generates a value of
59.9, TSL-LRSR [28] generates a value of 61.9, auto encoder [31] generates a value of 62.1, JDA [25]
generates a value of 64.8. CORAL [24] generates a value of 64.9, TCA [26] generates a value of 65.7, DBN
[32] generates a value of 68, DiagNet [33] generates a value of 68.8, MEDA generates a value of 69, CMMS
[29] generates a value of 70, Fed_Align generates a value of 71.1, LRCDR [ES-existing system] generates a
value of 73.1. Whereas PS depicts a value of 83.46 which gives better performance in comparison with
existing system.
Table 1 Accuracy comparison
Method Accuracy
maLRR 59.9
TSL-LRSR 61.9
Auto Encoder 62.1
JDA 64.8
CORAL 64.9
TCA 65.7
DBN 68
DiagNet 68.8
MEDA 69
CMMS 70
Fed_Align 71.1
LRCDR [ES] 73.1
Proposed System (PS) 83.46

2151
Figure 4 shows the sensitivity metric comparison is carried out with various existing state-of-art
methods. Where TSL-LRSR [28] generates a value of 36.4, maLRR [30] generates a value of 52.9, JDA [25]
generates a value of 57.7, CORAL [24] generates a value of 59.7, Fed_Align generates a value of 64.3, Auto
encoder [31] generates a value of 65.6, MEDA generates a valu e of 66.5. DiagNet [33] generates a value of
68.8TCA [26] generates a value of 66.9, CMMS [29] generates a value of 68, DBN [32] generates a value of
68.4, LRCDR [ES] generates a value of 71 whereas PS depicts a value of 81.24, which gives better
performance in comparison with existing system. Table 2 shows the sensitivity comparison.
Figure 3. Accuracy comparison for various existing
methods in comparison with proposed system
Figure 4. Sensitivity comparison for various existing
Table 2. Sensitivity comparison
Method Sensitivity
TSL-LRSR 36.4
maLRR 52.9
JDA 57.7
CORAL 59.7
Fed_Align 64.3
Autoencoder 65.6
MEDA 66.5
DiagNet 66.6
TCA 66.9
CMMS 68
DBN 68.4
LRCDR 71
PS 81.24
Figure 5 shows the specificity metric comparison is carried out with various existing state-of-art
methods. Table 3 shows the specificity comparison. Table 4 show the F1-score comparison. Figure 6 shows
the F1-score comparison graph. Where TSL-LRSR [28] generates a value of 36.4, maLRR [30] generates a
value of 52.9, JDA [25] generates a value of 57.7, CORAL [24] generates a value of 59.7. Fed Align
generates a value of 64.3, auto encoder [31] generates a value of 65.6, MEDA generates a valu e of 66.5,
DiagNet [33] generates a value of 68.8TCA [26] generates a value of 66.9, CMMS [29] generates a value of
68. DBN [32] generates a value of 68.4, LRCDR [ES] generates a value of 71 whereas PS depicts a value of
81.24. Proposed system gives better performance in comparison with existing system.
2.4. Comparative analysis
The comparative analysis is carried out further in comparison with the existing system and proposed
system for various performance metric is evaluated and the improvisation in % is shown in the Table 5. For
Accuracy metric the ES generates a value of 73.1 and PS generates a value of 78.46 and the improvisation is
7.07311%. whereas for sensitivity metric the ES generates a value of 71 and PS generates a value of 81.24
and the improvisation is 13.5424%. For specificity metric the ES generates a value of 75.1, PS generates a
value of 87.43%, and the improvisation is 15.1726% whereas for F-score metric, the ES generates a value of
73.8 and PS generates a value of 83.8 and the improvisation is 12.69045%.

 ISSN: 2252-8938
2152
Figure 5 Specificity comparison for various existing
Figure 6. F1-score comparison for various existing
Table 3. Specificity comparision
Method Specificity
TCA 65.1
maLRR 66.3
Autoencoder 67.3
DBN 68.5
CORAL 69.8
DiagNet 71
JDA 71.4
MEDA 71.4
CMMS 72
Fed_Align 77.7
LRCDR [ES] 75.1
TSL-LRSR 86.2
PS 87.43
Table 4. F1-Score comparison
Method F1
Autoencoder 55.6
maLRR 62.5
DBN 64.4
TCA 64.6
CORAL 67.1
Fed_Align 67.8
JDA 68.3
DiagNet 68.7
CMMS 70.7
LRCDR [ES] 73.8
TSL-LRSR 79
PS 83.8
Table 5 Comparison table
Performance metric ES PS Improvisation (%)
Accuracy 73.1 78.46 7.07311
Sensitivity 71 81.24 13.4524
Specificity 75.1 87.43 15.1726
F-score 73.8 83.8 12.69045
4. CONCLUSION
A deep learning model called the attribute feature graph layer has been developed for identifying
individuals with ASD. The model utilizes an attribute feature graph for identification purposes. This graph is
constructed by leveraging fMRI data to establish node characteristics. The learning process for node features
involves establishing connections within the attribute feature graph between nodes and their corresponding
attributes. The probability of ASD is determined by performing a weighted sum of the node features, which
have been processed through the attribute feature network. Based on the analysis of brain functional

2153
networks, this study proposes a framework for the classification of ASD. The framework integrates a
customized CNN with experiential learning strategies. Transfer learning techniques are employed to enhance
the convergence speed of models and optimize classification performance. The proposed model demonstrates
improved performance in categorizing functional brain networks on large-scale datasets, as evidenced by
rigorous testing conducted on the ABIDE dataset. Based on the conducted comparative analysis, the accuracy
metric shows an improvement of 7.07311%, the sensitivity metric demonstrates an improvement of
13.4524%, the specificity metric exhibits an improvement of 15.1726%, and the F-score metric indicates an
improvement of 12.69045%. These results suggest that our proposed model performs effectively when
compared to the current system.
REFERENCES
[1] F. Ke, S. Choi, Y. H. Kang, K.-A. Cheon, and S. W. Lee, “Exploring the structural and strategic bases of autism spectrum
disorders with deep learning,” IEEE Access, vol. 8, pp. 153341–153352, 2020, doi: 10.1109/ACCESS.2020.3016734.
[2] Z. Lu, J. Wang, R. Mao, M. Lu, and J. Shi, “Jointly composite feature learning and autism spectrum disorder classification using
deep multi-output takagi-sugeno-kang fuzzy inference systems,” IEEE/ACM Transactions on Computational Biology and
Bioinformatics, vol. 20, no. 1, pp. 476-488, 2023, doi: 10.1109/TCBB.2022.3163140.
[3] Z. A. Huang, Z. Zhu, C. H. Yau, and K. C. Tan, “Identifying autism spectrum disorder from resting-state fMRI using deep belief
network,” IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 7, pp. 2847–2861, 2021, doi:
10.1109/TNNLS.2020.3007943.
[4] Z.-A. Huang et al., “Federated multi-task learning for joint diagnosis of multiple mental disorders on MRI scans,” IEEE
Transactions on Biomedical Engineering, vol. 70, no. 4, pp. 1137–1149, Apr. 2023, doi: 10.1109/TBME.2022.3210940.
[5] M. Eni, I. Dinstein, M. Ilan, I. Menashe, G. Meiri, and Y. Zigel, “Estimating autism severity in young children from speech
signals using a deep neural network,” IEEE Access, vol. 8, pp. 139489–139500, 2020, doi: 10.1109/ACCESS.2020.3012532.
[6] J. Han, G. Jiang, G. Ouyang, and X. Li, “A multimodal approach for identifying autism spectrum disorders in children,” IEEE
Transactions on Neural Systems and Rehabilitation Engineering, vol. 30, pp. 2003–2011, 2022, doi:
10.1109/TNSRE.2022.3192431.
[7] S. Liang, A. Q. M. Sabri, F. Alnajjar, and C. K. Loo, “Autism spectrum self-stimulatory behaviors classification using explainable
temporal coherency deep features and SVM classifier,” IEEE Access, vol. 9, pp. 34264–34275, 2021, doi:
10.1109/ACCESS.2021.3061455.
[8] F. Ke, S. Choi, Y. H. Kang, K.-A. Cheon, and S. W. Lee, “Exploring the structural and strategic bases of autism spectrum
disorders with deep learning,” IEEE Access, vol. 8, pp. 153341–153352, 2020, doi: 10.1109/ACCESS.2020.3016734.
[9] M. R. Ahmed, Y. Zhang, Y. Liu, and H. Liao, “Single volume image generator and deep learning-based ASD classification,”
IEEE Journal of Biomedical and Health Informatics, vol. 24, no. 11, pp. 3044–3054, 2020, doi: 10.1109/JBHI.2020.2998603.
[10] F. Ke and R. Yang, “Classification and biomarker exploration of autism spectrum disorders based on recurrent attention model,”
IEEE Access, vol. 8, pp. 216298–216307, 2020, doi: 10.1109/ACCESS.2020.3038479.
[11] C. Xia et al., “Dynamic viewing pattern analysis: Towards large-scale screening of children with ASD in remote areas,” IEEE
Transactions on Biomedical Engineering, vol. 70, no. 5, pp. 1622–1633, May 2023, doi: 10.1109/TBME.2022.3223736.
[12] M. N. A. Tawhid, S. Siuly, K. Wang, and H. Wang, “Automatic and efficient framework for identifying multiple neurological
disorders from EEG signals,” IEEE Transactions on Technology and Society, vol. 4, no. 1, pp. 76–86, Mar. 2023, doi:
10.1109/TTS.2023.3239526.
[13] Y. Zhang et al., “Predicting the Symptom Severity in Autism Spectrum Disorder Based on EEG Metrics,” IEEE Transactions on
Neural Systems and Rehabilitation Engineering, vol. 30, pp. 1898-1907, 2022, doi: 10.1109/TNSRE.2022.3188564.
[14] C. Han et al., “GAN-based multiple adjacent brain MRI slice reconstruction for unsupervised Alzheimer’s disease diagnosis,” in
Computational Intelligence Methods for Bioinformatics and Biostatistics, Cham: Springer, 2020, pp. 44–54, doi: 10.1007/978-3-
030-63061-4_5.
[15] M. Kohli, A. K. Kar, and S. Sinha, “The role of intelligent technologies in early detection of autism spectrum disorder (ASD): A
scoping review,” IEEE Access, vol. 10, pp. 104887–104913, 2022, doi: 10.1109/ACCESS.2022.3208587.
[16] G. Wen et al., “Graph self-supervised learning with application to brain networks analysis,” IEEE Journal of Biomedical and
Health Informatics, vol. 27, no. 8, pp. 4154–4165, 2023, doi: 10.1109/JBHI.2023.3274531.
[17] Y. Tang, D. Chen, H. Liu, C. Cai, and X. Li, “Deep EEG superresolution via correlating brain structural and functional
connectivities,” IEEE Transactions on Cybernetics, vol. 53, no. 7, pp. 4410–4422, Jul. 2023, doi: 10.1109/TCYB.2022.3178370.
[18] X. Li, N. C. Dvornek, J. Zhuang, P. Ventola, and J. S. Duncan, “Brain biomarker interpretation in ASD using deep learning and
fMRI,” in Medical Image Computing and Computer Assisted Intervention-MICCAI 2018, Cham: Springer, 2018, pp. 206–214,
doi: 10.1007/978-3-030-00931-1_24.
[19] C. Han et al., “MADGAN: unsupervised medical anomaly detection GAN using multiple adjacent brain MRI slice
reconstruction,” BMC Bioinformatics, vol. 22, pp. 1-20, 2021, doi: 10.1186/s12859-020-03936-1.
[20] L. Sun, J. Wang, Y. Huang, X. Ding, H. Greenspan, and J. Paisley, “An adversarial learning approach to medical image synthesis
for lesion detection,” IEEE Journal of Biomedical and Health Informatics, vol. 24, no. 8, pp. 2303–2314, Aug. 2020, doi:
10.1109/JBHI.2020.2964016.
[21] Q. Yao and H. Lu, “Brain functional connectivity augmentation method for mental disease classification with generative
adversarial network,” in Pattern Recognition and Computer Vision, Cham, Springer, 2019, pp. 444–455, doi: 10.1007/978-3-030-
31654-9_38.
[22] G. Nie et al., “An Immersive Computer-Mediated Caregiver-Child Interaction System for Young Children with Autism Spectrum
Disorder,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 29, pp. 884-893, 2021, doi:
10.1109/TNSRE.2021.3077480.
[23] A. Sadiq, M. I. Al-Hiyali, N. Yahya, T. B. Tang, and D. M. Khan, “Non-oscillatory connectivity approach for classification of
autism spectrum disorder subtypes using resting-state fMRI,” IEEE Access, vol. 10, pp. 14049–14061, 2022, doi:
10.1109/ACCESS.2022.3146719.
[24] B. Sun, J. Feng, and K. Saenko, “Return of frustratingly easy domain adaptation,” Proceedings of the AAAI Conference on
Artificial Intelligence, vol. 30, no. 1, Mar. 2016, doi: 10.1609/aaai.v30i1.10306.

 ISSN: 2252-8938
2154
[25] M. Long, J. Wang, G. Ding, J. Sun, and P. S. Yu, “Transfer feature learning with joint distribution adaptation,” in 2013 IEEE
International Conference on Computer Vision, Dec. 2013, pp. 2200–2207, doi: 10.1109/ICCV.2013.274.
[26] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain adaptation via transfer component analysis,” IEEE Transactions on
Neural Networks, vol. 22, no. 2, pp. 199–210, Feb. 2011, doi: 10.1109/TNN.2010.2091281.
[27] J. Wang, W. Feng, Y. Chen, H. Yu, M. Huang, and P. S. Yu, “Visual domain adaptation with manifold embedded distribution
alignment,” in Proceedings of the 26th ACM international conference on Multimedia, Oct. 2018, pp. 402–410, doi:
10.1145/3240508.3240512.
[28] Y. Xu, X. Fang, J. Wu, X. Li, and D. Zhang, “Discriminative transfer subspace learning via low-rank and sparse representation,”
IEEE Transactions on Image Processing, vol. 25, no. 2, pp. 850–863, Feb. 2016, doi: 10.1109/TIP.2015.2510498.
[29] L. Tian, Y. Tang, L. Hu, Z. Ren, and W. Zhang, “Domain adaptation by class centroid matching and local manifold self-learning,”
IEEE Transactions on Image Processing, vol. 29, pp. 9703–9718, 2020, doi: 10.1109/TIP.2020.3031220.
[30] M. Wang, D. Zhang, J. Huang, P. T. Yap, D. Shen, and M. Liu, “Identifying autism spectrum disorder with multi-site fMRI via
low-rank domain adaptation,” IEEE Transactions on Medical Imaging, vol. 39, no. 3, pp. 644–655, 2020, doi:
10.1109/TMI.2019.2933160.
[31] A. S. Heinsfeld, A. R. Franco, R. C. Craddock, A. Buchweitz, and F. Meneguzzi, “Identification of autism spectrum disorder
using deep learning and the ABIDE dataset,” NeuroImage: Clinical, vol. 17, pp. 16–23, 2018, doi: 10.1016/j.nicl.2017.08.017.
[32] Z. A. Huang, Z. Zhu, C. H. Yau, and K. C. Tan, “Identifying autism spectrum disorder from resting-state fMRI using deep belief
network,” IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 7, pp. 2847–2861, 2021, doi:
10.1109/TNNLS.2020.3007943.
[33] A. Fong, T. Eslami, F. Saeed, V. Mirjalili, and A. R. Laird, “ASD-DiagNet: A hybrid learning approach for detection of autism
spectrum disorder using fMRI data,” Frontiers in Neuroinformatics, vol. 13, pp. 1-11, 2019, doi: 10.3389/fninf.2019.00070.
[34] X. Li, Y. Gu, N. Dvornek, L. H. Staib, P. Ventola, and J. S. Duncan, “Multi-site fMRI analysis using privacy-preserving federated
learning and domain adaptation: ABIDE results,” Medical Image Analysis, vol. 65, pp. 1-14, 2020, doi:
10.1016/j.media.2020.101765.
BIOGRAPHIES OF AUTHOR
Shabeena Lylath is currently working as assistant professor in HKBK College of
Engineering, Bangalore, India, she earned his Bachelor of Engineering BE degree in ISE from
Channabasaveshwara Institute of Technology, Tumkur, India in 2005. She obtained her
master’s degree in M.Tech. (CSE) from Akshaya Institute of Technology, Tumkur, India in
2015. Currently she is a research scholar at REVA University doing her Ph.D. in School of
Computer Science and Engineering. She has an Experience of 10 years in teaching in
Computer Science and Engineering. Her areas of interest are artificial intelligence and machine
learning. She can be contacted at email: shabeenal_12@rediffmail.com.
Laxmi B. Rananavare is currently working as associate professor School of
Computing and Information Technology, REVA University, Bangalore, India. She has vast
experience of more than 30 years in teaching in Computer Science and Engineering. Her
research interests are: artificial intelligence, machine learning, and natural language
processing. She is involved in research of applications of artificial intelligence. She has
published many in national and international conferences and journals. She can be contacted at
email: laxmib.rananavare@reva.edu.in.

Autism spectrum disorder identification with multi-site functional magnetic resonance imaging

More Related Content

Similar to Autism spectrum disorder identification with multi-site functional magnetic resonance imaging (20)

More from IAESIJAI (20)

Recently uploaded (20)

Autism spectrum disorder identification with multi-site functional magnetic resonance imaging