Computers in Biology and Medicine 182 (2024) 109083

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/compbiomed

MADE-for-ASD: A multi-atlas deep ensemble network for diagnosing Autism

Spectrum Disorder
Xuehan Liu a ,1 , Md Rakibul Hasan b,d ,1 , Tom Gedeon b,c , Md Zakir Hossain a,b ,∗
a
Australian National University, Canberra, ACT, 2601, Australia
b
Curtin University, Bentley, WA, 6102, Australia
c
Obuda University, Budapest, 1034, Hungary
d
BRAC University, Dhaka, 1212, Bangladesh

ARTICLE INFO ABSTRACT

Keywords: In response to the global need for efficient early diagnosis of Autism Spectrum Disorder (ASD), this paper
Autism bridges the gap between traditional, time-consuming diagnostic methods and potential automated solutions.
Neuroimaging We propose a multi-atlas deep ensemble network, MADE-for-ASD, that integrates multiple atlases of the brain’s
Computer vision
functional magnetic resonance imaging (fMRI) data through a weighted deep ensemble network. Our approach
Deep learning
integrates demographic information into the prediction workflow, which enhances ASD diagnosis performance
Health computing
and offers a more holistic perspective on patient profiling. We experiment with the well-known publicly
available ABIDE (Autism Brain Imaging Data Exchange) I dataset, consisting of resting state fMRI data from 17
different laboratories around the globe. Our proposed system achieves 75.20% accuracy on the entire dataset
and 96.40% on a specific subset — both surpassing reported ASD diagnosis accuracy in ABIDE I fMRI studies.
Specifically, our model improves by 4.4 percentage points over prior works on the same amount of data.
The model exhibits a sensitivity of 82.90% and a specificity of 69.70% on the entire dataset, and 91.00%
and 99.50%, respectively, on the specific subset. We leverage the F-score to pinpoint the top 10 ROI in ASD
diagnosis, such as precuneus and anterior cingulate/ventromedial. The proposed system can potentially pave
the way for more cost-effective, efficient and scalable strategies in ASD diagnosis. Codes and evaluations are
publicly available at https://github.com/hasan-rakibul/MADE-for-ASD.

1. Introduction Observation Schedule (ADOS) [6] and the Autism Diagnostic Interview-
Revised (ADI-R) [7]. Both ADOS and ADI-R deliver important insights
Autism Spectrum Disorder (ASD) is a prevalent neurodevelopmental into an individual’s communication, social interactions and behaviour.
condition characterised by challenges in social and communicative
Nevertheless, these methods primarily depend on observing symptoms,
abilities, as well as repetitive and hard-to-control behaviours in daily
making them largely subjective. The process can be both expensive
life [1]. ASD often co-occurs with various other conditions, such as in-
tellectual impairment, seizures and anxiety. Autistic individuals display and lengthy. Additionally, variations in symptom presentations and
a wide range of characteristics, from mild to severe social and commu- their severity can introduce additional complications, underscoring the
nicative differences, along with restricted and repetitive behaviours and need for a diagnostic method that is more efficient and objective [8].
interests [2]. According to a 2022 report endorsed by the World Health In this context, functional magnetic resonance imaging (fMRI) of the
Organisation2 [3], approximately one in 100 children worldwide are brain could be leveraged because it provides a non-invasive means
autistic, a significant increase from the one in 160 reported in 2012 [4]. of examining brain activity. It has the potential to elucidate the in-
The economic impact on families of autistic individuals is substantial,
tricate neurological deviations representing ASD, which can be used
making ASD a critical public health concern [5].
Diagnosis of ASD at present leverages a variety of techniques, towards developing automated, objective, efficient and early diagnostic
predominantly behavioural assessments such as the Autism Diagnostic methods [9,10].

∗ Corresponding author at: Australian National University, Canberra, ACT, 2601, Australia.
E-mail addresses: u7094891@alumni.anu.edu.au (X. Liu), Rakibul.Hasan@curtin.edu.au (M.R. Hasan), Tom.Gedeon@curtin.edu.au (T. Gedeon),
Zakir.Hossain1@curtin.edu.au (M.Z. Hossain).
1
Xuehan Liu and Md Rakibul Hasan contributed equally to this work.
2
https://www.who.int/news-room/fact-sheets/detail/autism-spectrum-disorders

https://doi.org/10.1016/j.compbiomed.2024.109083
Received 11 February 2024; Received in revised form 22 August 2024; Accepted 28 August 2024
Available online 3 September 2024
0010-4825/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

The Autism Brain Imaging Data Exchange (ABIDE) consortium re- diagnosis have leveraged a variety of machine learning (ML) and deep
leases resting-state fMRI (rs-fMRI) data with T1 structural brain images learning (DL) models with rs-fMRI data from the ABIDE datasets [9,12–
and demographic information of autistic individuals and Typical Con- 22]. For example, Abraham et al. [12] experimented with several
trols (TCs) [11]. ABIDE I dataset was collected from 17 international ML algorithms, including support vector regression and ridge regres-
research sites, which makes it diverse and comprehensive for under- sion, achieving a classification accuracy of 66.8%. However, these
standing the neurological nuances of ASD. Such heterogeneity in the conventional machine learning models have limited ability to learn
data can help capture the diverse manifestations of ASD across different complex patterns from raw data and do not take advantage of deep
populations and geographical locations. ABIDE I data, therefore, can representations. In a separate study, Heinsfeld et al. [13] embraced a
strengthen the generalisability of findings and enhance their clinical deep learning technique with an autoencoder (AE) and deep neural
relevance and applicability. The preference for using ABIDE I data is network (DNN), reaching 70% accuracy. Eslami et al. [19] leveraged
proven by many prior works on ASD diagnosis [9,12–22]. This paper an AE with a single-layer perceptron and achieved 70.3% accuracy,
uses ABIDE I dataset to classify between ASD and TC, hereinafter while Almuqhim and Saeed [20] used an AE, achieving 70.8% clas-
referred to as ASD diagnosis. sification accuracy. Although these studies have shown the proof of
Machine learning (ML) and deep learning (DL) techniques have concept of using AEs in this task, their performance remains limited.
been increasingly employed to advance the understanding and diag- Few studies [27–29] utilised a subset of the ABIDE dataset. For
nosis of various neurodevelopmental disorders, including ASD. ML example, Plitt et al. [29] employed a random forest classifier to predict
and DL provide an objective and data-driven approach to diagnosis, ASD on only 179 samples and reported a classification accuracy of
thereby reducing the reliance on subjective symptom-based criteria. 95%. Chen et al. [28] selected 252 high-quality data based on criteria
However, building appropriate models with high-dimensional multi- such as head motion, artifacts, and signal dropout and reported an
site neuroimaging data, such as ABIDE I, is challenging because of accuracy of 91% using a random forest classifier.
added dimensionality (i.e., different brain regions may have different Some studies [16,30] used both ABIDE I and II datasets. For ex-
cues towards ASD diagnosis) and site diversity (i.e., different data col- ample, Khosla et al. [16] trained their Convolutional Neural Network
lection sites have different data collection protocols). Including other (CNN) architecture with the ABIDE-I data and tested their model’s
important cues, such as people’s demographic information, can po- performance on the ABIDE-II dataset. Aghdam et al. [30] proposed a
tentially improve the predictive performance but make the system ‘mixture of experts’ ensemble approach to diagnosing autistic young
development more challenging while accommodating these data with children (aged 5–10 years) in three dataset conditions: ABIDE I, ABIDE
the primary neuroimaging data. Addressing these challenges, we pro- II, and a combination of them.
pose a novel DL-based ASD diagnosis workflow from ABIDE I brain The National Database for Autism Research (NDAR) is another
fMRI data and corresponding demographic information. The input significant repository, offering a wide range of data types, includ-
rs-fMRI images are processed to extract regions of interest (ROIs)
ing neuroimaging data, collected from various studies and institu-
according to three different atlases (brain parcellations), from where
tions [31]. Using this dataset, Li et al. [32] proposed a multi-channel
functional connectivity features are extracted. This paper’s novelty is
CNN model for early ASD diagnosis. Despite the variety of neuroimag-
two-fold. (1) It proposes MADE-for-ASD, consisting of a stacked sparse
ing datasets available, the ABIDE I dataset remains the most widely
denoising autoencoder (SSDAE) and multi-layer perceptron (MLP), fol-
used in ML-based ASD diagnosis [33].
lowed by a weighted ensemble learning framework. We incorporate
demographic information into the model, which enhances the per-
2.2. Selection of brain atlases
formance through personalised prediction. (2) This paper provides
visualised insights into the most significant ROIs with a high cor-
relation with ASD, which could further assist in understanding the Atlas-based parcellation, which divides the entire brain into spa-
neurobiological underpinnings of ASD. tially proximate ROIs, offers several advantages in neuroimaging stud-
ies [34]: it identifies brain regions with significant connectivity differ-
2. Background and related work ences between groups, examines brain functional organisation, reduces
data dimensionality and improves result interpretability by linking
2.1. Neuroimaging in ASD diagnosis specific brain regions to conditions or phenotypes. The Automated
Anatomical Labelling (AAL) atlas, with 116 ROIs, is widely used in
Magnetic resonance imaging (MRI) is a critical tool for understand- ASD diagnosis studies [16,35–37] due to its ability to provide pre-
ing the pathophysiology of neurological disorders such as schizophrenia cise anatomical locations essential for identifying structural abnormal-
and autism. MRI is valuable due to its cost-effectiveness and non- ities caused by ASD. The Craddock 200 atlas [38], with 200 ROIs, is
invasive nature, which have led to its widespread acceptance and based on functional connectivity and is particularly useful for analysing
application in the medical community [23]. Specifically, functional resting-state fMRI data [9,36,37,39], as it clusters the brain into func-
MRI (fMRI) tracks changes in blood oxygen levels over time, making tionally homogeneous regions.
it adept at inferring brain activity and drawing considerable attention Recent studies have leveraged multiple atlases to capture diverse
from researchers studying brain dysfunctions [24]. Changes in the patterns related to ASD. For instance, Mahler et al. [37] proposed a
intensity of fMRI images throughout the acquisition period usually multi-atlas framework for ASD classification using resting-state fMRI
serve as a representation of brain activity, typically expressed as a data, employing three atlases, including AAL and Craddock 200. Simi-
time series. Brain disorders rarely manifest as anomalies in singular or larly, Deng et al. [9] utilised the AAL, Craddock 200 and Craddock 400
multiple brain regions; they typically manifest as atypical connectivity atlases. Their experiment indicates that Craddock 200 can outperform
among various brain regions. In this context, functional connectivity Craddock 400 in fMRI-based ASD diagnosis. Hence, we selected the
helps investigate the association of specific activities between brain Craddock 200 atlas, hereinafter referred to as the CC atlas.
regions and has found widespread use in the classification of brain To further enhance the analysis, we leverage the Eickhoff–Zilles
disorders [25]. (EZ) atlas with 116 ROIs, which has also been utilised in recent ASD
Several datasets are instrumental in advancing autism research by diagnosis studies [40,41]. Derived from cytoarchitectonic mapping,
providing extensive neuroimaging data. The ABIDE I dataset aggre- the EZ atlas combines anatomical and functional data, offering a nu-
gates resting-state fMRI data from 17 international sites [11]. ABIDE anced view that bridges structural and functional aspects of brain
II expands on this by including additional subjects and sites, further regions [42]. By integrating the complementary strengths of AAL, CC
enhancing its utility for ASD research [26]. Various works in ASD and EZ atlases to capture different aspects of brain structure and

2
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

Fig. 1. Overall framework of our ASD/TC classification workflow. We first calculate functional connectivity matrices on three brain atlases, followed by feature selection using
F-score. To diagnose ASD, we use deep learning with a stacked sparse denoising autoencoder (SSDAE) and multi-layer perceptron (MLP), followed by a weighted ensemble. In the
case of extracting high-quality data, we train the classifier with the NYU subset and predict on the whole dataset.

function, our approach ensures a comprehensive analysis crucial for ASD and TC subjects is functional connectivity. We compute Pearson
accurate ASD diagnosis. correlation coefficients between the time series of each pair of brain
In a typical fMRI-based ASD diagnosis pipeline [9,16], atlas data regions to produce a connectivity matrix:
is converted to functional connectivity matrices, which measures the 𝐸(𝑢𝑣) − 𝐸(𝑢)𝐸(𝑣)
degree of synchronised activity between different brain regions based PCC (𝑢, 𝑣) = √ ( ) √ ( ) (1)
on the time series of resting-state fMRI brain imaging data. While 𝐸 𝑢2 − 𝐸 2 (𝑢) 𝐸 𝑣2 − 𝐸 2 (𝑣)
two of the atlases (AAL and EZ) focus on anatomical locations, they where PCC (𝑢, 𝑣) stands for the Pearson correlation coefficient between
also provide a foundational framework for understanding the structural time series of two brain regions 𝑢 and 𝑣, and 𝐸(⋅) is the mathematical
context within which functional connectivity occurs. Accordingly, func- expectation. These coefficients range from −1 to 1; coefficients near
tional connectivity of AAL and EZ atlases are utilised in ASD diagnosis 1 indicate a strong positive correlation, while those near −1 indicate
literature [9,16,40]. a strong negative correlation between the time series of two brain
regions. For example, we obtain a 200 × 200 symmetric matrix for
3. Method correlation in the case of the CC atlas because it has been divided into
200 regions. This functional connectivity matrix is used as a feature to
3.1. Task formulation classify subjects into ASD and TC groups.
To estimate the duplicated values in the matrix, we take the upper
Let us consider 𝑋 = {𝑋fMRI , 𝑋demog }, where 𝑋fMRI and 𝑋demog triangle of the symmetric matrix as the original feature representation
are the input fMRI images and demographic/phenotypic information, of this subject. We then flatten the remaining triangle by collapsing it
respectively, and 𝑌 = {𝑌1 , 𝑌2 }, where 𝑌1 and 𝑌2 are ASD and TC classes, in a one-dimension vector to retrieve a vector of features:
respectively. The goal is to develop a binary classifier  to predict (𝑁 − 1)𝑁
𝑆= (2)
whether a sample input 𝑥𝑖 ∈ 𝑋 can be classified as 𝑦𝑖 ∈ 𝑌 . 2
where 𝑆 is the dimension for the flattened vector, and 𝑁 is the number
3.2. Overall framework of the regions in the atlas. For example, we get a 1-D vector with
19,900 features for each sample for the CC atlas. Following the previous
Fig. 1 illustrates the overall framework of our ASD diagnosis system, computational process, for every individual subject, we secure three
which is fundamentally segmented into two distinct stages: the ASD/TC functional connectivity feature representations based on the respective
classification phase and the high-quality subset selection phase. three atlases.

3.4. Feature selection

3.3. Data preprocessing

Using F-score [45], we rank all features in descending order to

As data quality control, we exclude all samples with missing fMRI
prioritise those with the highest discriminative power between ASD
time series. The missing values in the demographic data of the samples
and TC subjects. Mathematically, the F-score is a ratio of variance
are imputed using the mean value of all available data of corresponding
between groups to variance within groups, quantifying each feature’s
categories [43]. In this way, the overall distribution of the data is
discriminatory power. A higher F-score indicates a larger difference
preserved, and the imputed values are likely to be close to the true in means relative to variability, suggesting better class distinction.
values, assuming that the missingness is completely random. While We compute the F-score values for all features in the dataset. Let 𝑥𝑘
we used the mean value to impute missing data, a more sophisticated represent the training vectors, with 𝑘 ranging from 1 to 𝑚, 𝑛+ being
approach, such as MICE [44], could be explored in future work. the count of positive instances, and 𝑛− the count of negative instances.
We extract the mean time series of ROIs for each sample. We The F-score for the 𝑖th feature is computed as follows:
use parcellated regions as our targeted ROIs to extract voxel-level ( )2 ( )2
connectivity features. For each atlas, a respective connectivity matrix 𝑟̄(+)
𝑖 − 𝑟̄𝑖 + 𝑟̄(−)
𝑖 − 𝑟̄𝑖
is formed, which is then condensed into a vector before being inputted 𝐹 (𝑖) = ( )2 ( )2 (3)
1 ∑ 𝑛+ 1 ∑𝑛−
into our model. The primary feature we use to differentiate between 𝑛 −1 𝑘=1
𝑟(+)
𝑘,𝑖
− 𝑟̄(+)
𝑖 + 𝑛−1 (−) (−)
𝑘=1 𝑟𝑘,𝑖 − 𝑟̄𝑖
+

3
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

Fig. 2. Overall architecture and training workflow of the deep networks of the MADE-for-ASD model. Knowledge from stacked sparse denoising autoencoders is transferred to a
multi-layer perceptron for ASD diagnosis. Here, 𝑋 refers to the input data, 𝑌 = {𝑌1 , 𝑌2 } are ASD and TC classes, and 𝑊 refers to the weight parameter.

Table 1 in the data, with sparsity constraints promoting sparse, distributed

Parameter configurations of the Autoencoders (AE1, AE2) and Multi-Layer Perceptron
representations.
(MLP) of the MADE-for-ASD model. Here, ‘n’ represents the number of input features.
The second component is a supervised learning stage using a Multi-
Parameter AE1 AE2 MLP
Layer Perceptron (MLP). The parameters learned from the SSDAE are
Structure n-1000-n 1000-600-1000 n-1000-500-104-2
transferred to the MLP’s first two layers, followed by fine-tuning to
Hidden layers 1 1 3
Learning rate 0.001 0.001 0.0005 enhance performance on the ASD classification task.
Batch size 100 10 10
Training iteration 700 1000 200 3.5.1. Stacked sparse denoising autoencoder
Optimiser GD GD SGD
Loss function Cross entropy Cross entropy Cross entropy
An Autoencoder (AE) consists of input, hidden and output lay-
Activation function tanh tanh tanh ers, with the hidden layer containing fewer neurons. Once feature
Momentum range / / 0.1-0.9 representation is obtained in an AE, it can be used to train a new
Dropout rate 0.5 0.5 0.3 AE, leading to a stacked AE with multiple layers. A Stacked Sparse
Sparsity parameter 0.5 0.5 /
Denoising Autoencoder (SSDAE) incorporates sparsity constraints and
Sparse penalty 0.2 0.2 /
Noise proportion 0.3 0.1 / noise addition to the input data for regularisation, enhancing model
robustness. The objective function includes a reconstruction loss term
and a sparsity penalty term, defined as:

1 ∑ ∑
𝑚 𝑠
where 𝑟̄𝑖 , 𝑟̄(+) (−)
𝑖 , and 𝑟̄𝑖 denote the average of the 𝑖th feature for the 𝐽 (𝛩) = 𝐿(𝑥(𝑖) , 𝑥̂ (𝑖) ) + 𝛽 KL(𝜌 ∥ 𝜌̂𝑗 ) (4)
complete dataset, the positive subset and the negative subset, respec- 𝑚 𝑖=1 𝑗=1
tively. The numerator captures the discriminatory power between the
where 𝐿(𝑥(𝑖) , 𝑥̂ (𝑖) ) denotes reconstruction loss, 𝛽 is the sparsity weight
positive and negative subsets, while the denominator encapsulates the
and KL(𝜌 ∥ 𝜌̂𝑗 ) is the Kullback–Leibler divergence between target spar-
dispersion within each subset.
sity 𝜌 and average activation 𝜌̂𝑗 of hidden unit 𝑗. The Kullback–Leibler
We retain the top 15% of ranked features. To determine this feature
divergence can be computed as:
retention range, we conducted a series of tests with selections ranging
from 5% to 50% based on different parcellations. The optimal param- 𝜌 1−𝜌
KL(𝜌 ∥ 𝜌̂𝑗 ) = 𝜌 log + (1 − 𝜌) log (5)
eters were determined based on the classifier’s average performance 𝜌̂𝑗 1 − 𝜌̂𝑗
across different parcellations. Details of this experiment are presented We employ an SSDAE with two hidden layers for unsupervised
in Appendix A. pre-training, as shown in the left part of Fig. 2. Optimal model perfor-
The dimensions of the post-preprocessing data vary across different mance on the validation set is achieved using reconstruction loss (mean
atlases; for example, features from the CC atlas have a dimensionality squared error). The input and output layers have 𝑁 features, where 𝑁
of 19,900, while the AAL and EZ atlases have a significantly smaller di- is the number of input features. The configurations and parameters are
mensionality of 6,670. To account for these differences and to maintain detailed in Table 1.
the 15% feature retention, we adopted an adaptive feature selection
approach. We retained the top 1,000 ranked features for the AAL and 3.5.2. Transfer learning to MLP
EZ atlases and the top 3,000 ranked features for the CC atlas. The SSDAE knowledge is transferred to an MLP with three hidden
layers. The first two layers, with 1,000 and 500 units, inherit the SSDAE
3.5. The MADE-for-ASD model parameters, while the third layer is initialised with random weights.
Four demographic features are added to the third layer as additional
The MADE-for-ASD model consists of two primary components, input. We add these into the last two layers because these features are
as illustrated in Fig. 2. The initial component is an unsupervised less significant if we add them into the original input compared to the
learning stage employing a Stacked Sparse Denoising Autoencoder large amounts of other features. The right section of Fig. 2 depicts the
(SSDAE). Each layer is trained independently, capturing key variations supervised training stage.

4
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

4.1.1. Preprocessing
Preprocessed ABIDE I data with four different pipelines are released
under Preprocessed Connectomes Project.3 [47] The number of data
varies across the resultant datasets from different pipelines. It is im-
portant to select a specific pipeline for a fair comparison with the
prior works. Most of the prior works [9,12–17,19–22] leveraged CPAC
Fig. 3. The multi-atlas weighted ensemble voting strategy. (Configurable Pipeline for the Analysis of Connectomes) pipeline. A
few studies, however, experimented with other pipelines, such as CCS
(Connectome Computation System) by Dvornek et al. [18] and DPARSF
Fine-tuning adjusts the MLP weights to minimise prediction error (Data Processing Assistant for Resting-State fMRI) by Mahler et al. [37].
in supervised tasks. The output layer consists of two units representing The CPAC pipeline includes several operations, such as slice timing
the likelihood of ASD or TC, using a softmax activation function to correction, motion correction and voxel intensity normalisation. In
normalise the output distribution and enable the outputs to represent addition, the nuisance signal was removed utilising 24 motion param-
the corresponding probabilities belonging to a particular class. The eters, CompCor with five components, low-frequency drifts (linear and
configurations of the MLP parameters are shown in Table 1. quadratic trends), and the global signal as regressors. Functional data
underwent band-pass filtering (0.01–0.1 Hz) and spatial registration
using a non-linear method to a template space (MNI152). We, therefore,
3.5.3. Weighed ensemble voting
leverage the CPAC pipeline, hereinafter referred to as the ABIDE I CPAC
Ensemble learning uses individual models and solves the same prob-
data.
lem. This work adopts the bagging ensemble approach, which involves
Although Di Martino et al. [11] released the ABIDE I dataset with
creating multiple subsets of the original data, training a model on each
1,112 samples, preprocessing and quality control reduced this number
and combining their predictions, often through majority voting, to form
to 1,035, which is consistent with prior studies [13,19,20,22]. For the
a final prediction [46].
NYU subset, our preprocessing and quality control reduced the number
Voting can be hard or soft. In hard voting, each classifier in the of samples from 182 to 175.
ensemble votes for one class label, and the class label that gets the
majority of votes is predicted. Our method is based on soft voting, as 4.1.2. Subset selection
shown in Fig. 3, which predicts based on the probabilities for each Previous research [27,29] demonstrated better performance with
class label. We assign a weight for each classifier according to their the NYU subset than the whole ABIDE I dataset. This suggests that the
individual classification accuracy among all three classifiers: NYU subset may possess lower noise levels and higher data quality.
acc Based on this, we hypothesise that using the NYU subset for training
𝑤𝑖 = ∑ 𝑖 (6)
acc𝑖 could improve the model’s ability to select high-quality data from the
entire ABIDE I dataset. We, therefore, train our model on the NYU
We then compute the sum of the products of weights and proba-
dataset using the CC atlas, save this model and repurpose it as a
bilities corresponding to each class across all classifiers. Subsequently,
selector. We apply this pretrained model to classify ASD/TC instances
the class with the highest cumulative value is designated as the output
across the whole ABIDE I dataset, which correctly classifies 645 out of
category:
1,035 samples (around 62.3%). We separate these correctly classified
∑ ( )
𝑦 = argmax𝑗∈𝑌 𝑤𝑖 ⋅ P 𝑖 (𝑥) = 𝑗 (7) samples to create a new subset, which potentially has higher quality.
( ) We refer to it as our proposed subset throughout this paper. Analyses on
where P 𝑖 (𝑥) = 𝑗 is the predicted probability that instance 𝑥 belongs the quality of our proposed subset compared to the whole dataset and
to class 𝑗 according to classifier 𝑖 . the NYU subset are presented in Appendix C.
As for the evaluation metrics, we report classification performance While this subset selection aims to improve data quality, it may
in terms of sensitivity and specificity in addition to accuracy because also lead to the selection of easier examples, potentially inflating per-
of the unbalanced nature of the dataset. formance metrics. Furthermore, this method may introduce exclusion
criteria that could affect the generalisability of the results. A thorough
4. Experiments analysis of the discarded subset to assess whether it excludes TC, ASD,
or both is warranted, which we leave for future work.
4.1. Dataset
4.2. ASD vs TC classification

We use rs-fMRI and demographic data from ABIDE I [11], the

We evaluate our MADE-for-ASD model through experiments on
first phase of ABIDE, with 505 autistic individuals and 530 Typical
different input data, ablation studies and comparative analyses with
Controls (TCs). We include four key demographic features of the sub-
state-of-the-art methods. Following the approach of prior ASD diagnosis
jects: age (years), sex (male/female), handedness (the dominant hand;
studies using the ABIDE I dataset [12–15,17,19–22], we employ 10-fold
left/ambiguous/right) and full-scale IQ (overall intellectual ability) (see
cross-validation to ensure robust evaluation and mitigate overfitting.
Tables B.6 and B.7 in Appendix B for the distribution of demographic This technique splits the dataset into ten equal-sized subsets, with
information across ASD and TC classes and the number of missing the model being trained on nine subsets and tested on the remaining
demographic data, respectively.). We also employ a subset of the T1w subset. This process is repeated ten times, and the classification scores
MRI images from the ABIDE I site, NYU Langone Medical Center, to are averaged across all runs, providing a cross-validated performance
test our proposed methodology. This subset encompasses 182 subjects, measure.
including 78 ASD and 104 TC subjects. We selected this specific subset
because it contains the largest amount of data among all participatory 4.2.1. Classification result
sites in ABIDE I. Additionally, previous studies [27,29] have reported Table 2 reports accuracy, sensitivity and specificity on the whole
ASD classification performance using the NYU subset, which allows us ABIDE I dataset, our proposed subset and the NYU subset.
to compare our model’s performance with theirs. Refer to Table B.8 in
Appendix B for details of the NYU subset having 78 autistic individuals
3
and 104 TC subjects. http://preprocessed-connectomes-project.org/abide/

5
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

Table 2
Classification performance of MADE-for-ASD model in terms of accuracy, sensitivity and specificity
on the whole ABIDE I data and its subsets.
Data # of Subjects Accuracy (%) Sensitivity (%) Specificity (%)
Whole dataset 1,035 75.20 82.90 69.70
Proposed subset 645 88.71 81.82 92.50
NYU subset 175 96.40 91.00 99.50

Table 3 further to 71.20% after removing the CC and EZ atlases. Removing the
Comparison with previous studies on whole ABIDE I CPAC data in terms of classification
CC and AAL atlases, the accuracy reaches the lowest at 68.74%. When
accuracy.
we only remove the EZ atlas, the accuracy is somewhat increased. This
Study Model # of Subjects Accuracy (%)
shows that integrating multiple atlases via voting contributes to the
Abraham et al. [12] SVC 871 66.80
improved performance of our model. The different atlas information
Heinsfeld et al. [13] AE+DNN 1,035 70.00
Sherkatghanad et al. [22] CNN 1,035 70.22 can complement each other, thereby enhancing the robustness of the
Eslami et al. [19] AE+SLP 1,035 70.30 classification task.
Parisot et al. [14] GCN 871 70.40 When we abolish the feature selection based on the F-score, it
Almuqhim and Saeed [20] SAE+MLP 1,035 70.80 leads to a decrease in accuracy to 72.80%. The drop in accuracy
Anirudh and Thiagarajan [21] Ensemble GCN 872 70.86
demonstrates the significance of feature selection. By focusing on the
Liu et al. [17] Extra-Trees 1,054 72.20
Wang et al. [15] Ensemble MLP 949 74.52 most informative features and deleting the noisy features, our model
Deng et al. [9] 3D-CNN 860 74.53 can better discriminate between classes. Lastly, removing the demo-
Ours (MADE-for-ASD) SSDAE+MLP 1,035 75.20 graphic information leads to an accuracy of 73.50%. This indicates that
including demographic information can aid in the classification process,
further improving the decision boundaries.
Although the inclusion of demographic data leads to better perfor-
The accuracy and specificity of the whole dataset are lower than
mance, these data should be carefully utilised with ML-based practical
those of other subsets, which can be attributed to the noise and com-
ASD diagnosis systems so that the system is not biased to particular
plexity in the whole dataset. The classifier reached its peak performance
demographics. Future work should focus on quantifying and mitigating
on the NYU subset, which can be attributed to the high quality of
such biases in ML models predicting ASD. Furthermore, exploring
NYU data. This is in line with Kong et al. [27] that the classification
performance is higher with the NYU subset. demographic differences within the pipeline – such as sex differences
in the top ROIs revealed by our algorithm – represents an important
4.2.2. Comparative analysis avenue for future research. Similarly, evaluating performance on strat-
We compare the performance of our model trained on the whole ified demographic subsets would help ensure that our model performs
ABIDE I dataset with the CPAC pipeline, which exhibits encouraging in- consistently across different groups.
sights ( Table 3). (1) Our model achieved an accuracy score of 75.20%, While our fMRI-based approach with deep learning provides ro-
which is a boost of 4.4 percentage points over the best prior work [20] bust diagnostic performance for ASD, it is important to acknowledge
on 1,035 samples. (2) Even without considering the use of the same the interpretability challenges associated with these methods. Tra-
amount of input data, our model outperforms the 74.53% accuracy on ditional behaviour-based approaches offer specific insights into the
860 samples reported by Deng et al. [9]. (3) By incorporating ensemble symptomatic nuances of participants, which can enhance the substance
techniques and introducing sparsity in the AE, our model outperforms of clinical reports. In contrast, our method, despite its diagnostic accu-
other AE and DNN-based approaches [13,19,20] by a margin of at least racy, may offer fewer insights into individual symptomatic specificities.
4.4 percentage points. This limitation is two-fold: first, there is a need for further research
We further compare our work with similar studies on certain subsets to improve the interpretation of fMRI data and its correlation with
of the ABIDE I dataset Table 4. Similar to the whole dataset, our autism symptoms; second, the deep learning model used in our study
proposed MADE-for-ASD model achieved the SOTA result on the NYU offers limited transparency, making it challenging to explain individ-
subset, showing an accuracy of 96.40%. Besides, our model’s perfor- ual predictions based on input features. Future work should enhance
mance on the new subset generated by our data selection achieved fMRI interpretability for autism to bridge the gap between diagnostic
an accuracy of 88.71%. This proves our model’s ability to perform accuracy and clinical insight.
consistently well on different data subsets.
This result also underlines the potential influence of data quality 4.3. Visualisation of top ROIs
on classification outcomes. The performance difference across subsets
reinforces the need for data selection in such classification tasks. Our We utilise the F-score feature selection methodology to discern and
strategy of creating a new subset from the ABIDE I dataset has also rank the most influential features. The initial 3D images from the CC
proven to be effective. The accuracy on this subset signifies the poten- atlas are clustered based on the spatial position of ROIs. These resulting
tial of using selective strategies for data preparation to boost model cluster centres are treated as the spatial coordinates of the ROIs. On
performance. an existing three-dimensional brain fMRI image from the dataset, we
then plot the spatial coordinates of these cluster centres. We rank the
4.2.3. Ablation study appearance frequencies of the associated ROIs, thereby identifying the
As can be seen on Table 5, we carry out a series of ablation exper- top 10 most frequently occurring areas of interest (Fig. 4).
iments to better understand the individual contributions of different As can be seen, the precuneus, typically recognised as a central node
components of the MADE-for-ASD model. These experiments include of the default mode network (DMN) [48], has a crucial impact on ASD
removing the ensemble voting and using single and combination of classification. This region is involved in self-referential thought and
atlas data for training, abolishing the F-score-based feature selection, social cognition, which are often disrupted in autistic individuals [49].
and omitting the sample demographic information during classification. Several studies have suggested that DMN connectivity can be associated
When we remove the voting and use only a single atlas and a combi- with a neurophenotype of ASD [16]. For example, Chen et al. [28] high-
nation of atlases, the classification accuracy decreases. When removing lighted substantial contributions from default mode and somatosensory
data of the AAL and EZ atlases, the accuracy drops to 73.42%. It drops areas towards ASD diagnosis. Similarly, Abraham et al. [12] found

6
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

Table 4
Comparison with previous studies on a subset of the ABIDE I dataset in terms of classification accuracy.
Study Data # of Subjects Model Accuracy (%)
Kong et al. [27] NYU 182 SAE 90.30
Chen et al. [28] Selected subseta 252 Random Forest 91.00
Plitt et al. [29] NYU, USM, UCLA 178 Random Forest 95.00
Ours (MADE-for-ASD) NYU 175 SSDAE+MLP 96.40
Ours (MADE-for-ASD) Proposed subset 645 SSDAE+MLP 88.71
a
Subset was selected based on specific criteria of high-quality data (e.g., low head motion).

Table 5
Classification accuracy in ablation study by removing atlases, feature selection and demographic
information. Negative sign (−) refers to the removal of the corresponding component.
Component Accuracy (%)
CC + AAL + EZ + Feature Selection + Demographic 75.20
− {AAL, EZ} 73.42
− {CC, EZ} 71.20
− {CC, AAL} 68.74
− EZ 74.10
− Feature Selection 72.80
− Demographic 73.50

Fig. 4. Top 10 most significant regions of interest towards ASD diagnosis using fMRI CC200 atlas.

distinguishing connections in the DMN related to ASD diagnosis using MLP, complemented by integrating demographic information, which
the ABIDE dataset. significantly enhances our model’s predictive capabilities. Furthermore,
The anterior cingulate/ventromedial prefrontal cortex, a region with our method of selecting high-quality classification subsets serves to
established connections to autism, was notably pronounced in the ASD reduce dataset noise and improve data quality. Our model achieves
classification problem [50]. This area is associated with emotional the SOTA accuracy on both the whole ABIDE I dataset (an improve-
regulation and decision-making, processes that are often impaired in ment of 4.4 percentage points) and its subset. Such an imaging-based
ASD [11]. Furthermore, anomalies in the medial prefrontal cortex ASD prediction system can benefit patients, families and healthcare
node of the DMN have been shown to detect social deficits in autistic systems worldwide through objective, efficient, non-intrusive and early
children [51]. Additionally, the left parietal cortex was stressed for ASD diagnosis.
prediction, aligning with the lateralised activation seen in this region
in autistic individuals [52]. These findings from the visualisation of
top ROIs corroborate previous studies that highlight the importance of CRediT authorship contribution statement
these regions in ASD pathology [16].

Xuehan Liu: Visualization, Validation, Software, Investigation, For-

5. Conclusion
mal analysis, Data curation. Md Rakibul Hasan: Writing – review
& editing, Writing – original draft, Visualization, Validation, Super-
ASD is a neurodevelopmental disorder characterised by a spec-
vision, Software, Project administration, Methodology, Investigation,
trum of symptoms and impairments. Common features of ASD include
Formal analysis. Tom Gedeon: Writing – review & editing, Supervi-
challenges with social interaction and communication, alongside a
preference for repetitive behaviours and interests, highlighting the sion, Resources, Methodology. Md Zakir Hossain: Writing – review
diverse nature of this condition. This paper proposes a novel ASD & editing, Validation, Supervision, Resources, Project administration,
diagnosis framework, MADE-for-ASD, involving a weighted ensemble Methodology, Funding acquisition, Conceptualization.
of DNNs using multi-atlas brain fMRI data. Through the F-score-based
feature selection method, we obtain discriminative features that offer
Declaration of competing interest
valuable visual insights into significant ROIs associated with ASD. They
shed light on the interplay of different features and their respective
contribution towards ASD diagnosis. This would help clinicians and The authors declare that they have no known competing finan-
researchers gain a more intuitive understanding of how different brain cial interests or personal relationships that could have appeared to
regions contribute to ASD. Our model consists of an SSDAE and an influence the work reported in this paper.

7
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

Fig. A.5. ASD classification accuracy with different feature selection ranges in our F-score-based feature selection. The presented accuracy refers to the average accuracy over all
three parcellations.

Table B.6
Distribution of demographic information of the ABIDE I subjects.
Site ASD TC
Age Sex Hand FIQ Age Sex Hand FIQ
CALTECH 27.4 (10.3) 15/4 0/5/14 108.2 (12.2) 28.0 (10.9) 14/4 1/3/14 114.8 (9.3)
CMU 26.4 (5.8) 11/3 1/1/12 114.5 (11.2) 26.8 (5.7) 10/3 0/1/12 114.6 (9.3)
KKI 10.0 (1.4) 16/4 1/3/16 97.9 (17.1) 10.0 (1.2) 20/8 1/3/24 112.1 (9.2)
LEUVEN 17.8 (5.0) 26/3 3/0/26 109.4 (12.6) 18.2 (5.1) 29/5 4/0/30 114.8 (12.4)
MAX_MUN 26.1 (14.9) 21/3 2/0/22 109.9 (14.2) 24.6 (8.8) 27/1 0/0/28 111.8 (9.1)
NYU 14.7 (7.1) 65/10 NaN 107.1 (16.3) 15.7 (6.2) 74/26 NaN 113.0 (13.3)
OHSU 11.4 (2.2) 12/0 1/0/11 106.0 (20.1) 10.1 (1.1) 14/0 0/0/14 115.0 (10.7)
OLIN 16.5 (3.4) 16/3 4/0/15 112.6 (17.8) 16.7 (3.6) 13/2 2/0/13 113.9 (16.0)
PITT 19.0 (7.3) 25/4 3/1/25 110.2 (14.3) 18.9 (6.6) 23/4 1/1/25 110.1 (9.2)
SBL 35.0 (10.4) 15/0 NaN NaN 33.7 (6.6) 15/0 NaN NaN
SDSU 14.7 (1.8) 13/1 1/0/13 111.4 (17.4) 14.2 (1.9) 16/6 3/0/19 108.1 (10.3)
STANFORD 10.0 (1.6) 15/4 3/1/15 110.7 (15.7) 10.0 (1.6) 16/4 0/2/18 112.1 (15.0)
TRINITY 16.8 (3.2) 22/0 0/0/22 108.9 (15.2) 17.1 (3.8) 25/0 0/0/25 112.5 (9.2)
UCLA 13.0 (2.5) 48/6 6/0/48 100.4 (13.4) 13.0 (1.9) 38/6 4/0/40 106.4 (11.1)
UM 13.2 (2.4) 57/9 7/8/51 105.5 (17.1) 14.8 (3.6) 56/18 9/2/63 108.2 (9.7)
USM 23.5 (8.3) 46/0 NaN 99.7 (16.4) 21.3 (8.4) 25/0 NaN 115.4 (14.8)
YALE 12.7 (3.0) 20/8 5/0/23 94.6 (21.2) 12.7 (2.8) 20/8 4/0/24 105.0 (17.1)

Age: Average (Standard Deviation), Sex: Male/Female, FIQ: Average (Standard Deviation).
Hand: Left/Ambiguous/Right; the dominant hand of samples.

Table B.7 Table B.8

Number of missing demographic data in the Class-wise statistics of the NYU subset.
ABIDE I dataset. Class Size Age Gender (F/M)
Demographic # of missing values
ASD 78 14.54 ± 5.29 10/68
information
TC 104 15.87 ± 5.04 26/78
Age 0
Sex 0
Handedness 326
Full-scale IQ 72 Appendix C. Subset quality analysis

Appendix A. Feature selection using F-score We analyse the age distribution in the NYU subset, our selected
subset and the whole dataset. As can be seen in Fig. C.6, samples
from younger individuals manifest a superior edge in the classification
Fig. A.5 illustrates our feature selection process based on F-score,
showing that the classification accuracy was highest when using the tasks. According to the higher classification accuracy on the NYU
top 15% of the total features. A histogram of F-scores for each atlas subset, the nature of these results suggests that age could play a
can reveal the decay and range of feature values, which should be pivotal role in shaping the outcomes of classification tasks. Given these
investigated in future work. findings, further exploration into the impact of age distribution on ASD
classification tasks may be discussed.
Appendix B. ABIDE I dataset statistics Our selected subset demonstrates a reduction in noise and vari-
ance compared to the whole dataset (Fig. C.7). This reduction poten-
The demographic details and sample distribution of the ABIDE I tially suggests higher-quality data because lower variance can enable
whole dataset and the NYU subset are presented in Table B.6 and the model to learn patterns more effectively with more accurate pre-
Table B.8, respectively. The number of missing demographic data in dictions. This aligns with findings that reduced data variability can
the ABIDE I dataset is illustrated in Table B.7. enhance model performance [53].

8
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

Fig. C.6. Demographic information distribution on different ranges of datasets.

Fig. C.7. Comparison of features variance between the whole dataset and (a) our proposed subset and (b) NYU subset.

References [9] J. Deng, M.R. Hasan, M. Mahmud, M.M. Hasan, K.A. Ahmed, M.Z. Hossain,
Diagnosing autism spectrum disorder using ensemble 3D-CNN: A preliminary
study, in: 2022 IEEE International Conference on Image Processing, ICIP, IEEE,
[1] D.G. Amaral, C.M. Schumann, C.W. Nordahl, Neuroanatomy of autism, Trends
2022, pp. 3480–3484, http://dx.doi.org/10.1109/ICIP46576.2022.9897628.
Neurosci. 31 (3) (2008) 137–145.
[10] T. Iidaka, Resting state functional magnetic resonance imaging and neural
[2] Y.S. Kim, B.L. Leventhal, Y.-J. Koh, E. Fombonne, E. Laska, E.-C. Lim, K.-A. network classified autism and control, Cortex 63 (2015) 55–67.
Cheon, S.-J. Kim, Y.-K. Kim, H. Lee, et al., Prevalence of autism spectrum [11] A. Di Martino, C.-G. Yan, Q. Li, E. Denio, F.X. Castellanos, K. Alaerts, J.S.
disorders in a total population sample, Am. J. Psychiatry 168 (9) (2011) Anderson, M. Assaf, S.Y. Bookheimer, M. Dapretto, et al., The autism brain
904–912. imaging data exchange: towards a large-scale evaluation of the intrinsic brain
[3] J. Zeidan, E. Fombonne, J. Scorah, A. Ibrahim, M.S. Durkin, S. Saxena, A. Yusuf, architecture in autism, Mol. Psychiatry 19 (6) (2014) 659–667.
A. Shih, M. Elsabbagh, Global prevalence of autism: A systematic review update, [12] A. Abraham, M.P. Milham, A. Di Martino, R.C. Craddock, D. Samaras, B. Thirion,
Autism Res. 15 (5) (2022) 778–790, http://dx.doi.org/10.1002/aur.2696. G. Varoquaux, Deriving reproducible biomarkers from multi-site resting-state
[4] M. Elsabbagh, G. Divan, Y.-J. Koh, Y.S. Kim, S. Kauchali, C. Marcín, C. Montiel- data: An Autism-based example, NeuroImage 147 (2017) 736–745, http://dx.
Nava, V. Patel, C.S. Paula, C. Wang, et al., Global prevalence of autism and doi.org/10.1016/j.neuroimage.2016.10.045.
other pervasive developmental disorders, Autism Res. 5 (3) (2012) 160–179, [13] A.S. Heinsfeld, A.R. Franco, R.C. Craddock, A. Buchweitz, F. Meneguzzi, Identi-
http://dx.doi.org/10.1002/aur.239. fication of autism spectrum disorder using deep learning and the ABIDE dataset,
[5] J.-J. Ou, L.-J. Shi, G.-L. Xun, C. Chen, R.-R. Wu, X.-R. Luo, F.-Y. Zhang, J.- NeuroImage: Clin. 17 (2018) 16–23, http://dx.doi.org/10.1016/j.nicl.2017.08.
P. Zhao, Employment and financial burden of families with preschool children 017.
[14] S. Parisot, S.I. Ktena, E. Ferrante, M. Lee, R. Guerrero, B. Glocker, D. Rueckert,
diagnosed with autism spectrum disorders in urban China: results from a
Disease prediction using graph convolutional networks: application to autism
descriptive study, BMC Psychiatry 15 (1) (2015) 1–8.
spectrum disorder and Alzheimer’s disease, Med. Image Anal. 48 (2018) 117–130,
[6] C. Lord, S. Risi, L. Lambrecht, E.H. Cook, B.L. Leventhal, P.C. DiLavore, A.
http://dx.doi.org/10.1016/j.media.2018.06.001.
Pickles, M. Rutter, The Autism Diagnostic Observation Schedule—Generic: A
[15] Y. Wang, J. Wang, F.-X. Wu, R. Hayrat, J. Liu, AIMAFE: Autism spectrum disorder
standard measure of social and communication deficits associated with the
identification with multi-atlas deep feature representation and ensemble learning,
spectrum of autism, J. Autism Dev. Disord. 30 (2000) 205–223.
J. Neurosci. Methods 343 (2020) 108840, http://dx.doi.org/10.1016/j.jneumeth.
[7] C. Lord, M. Rutter, A. Le Couteur, Autism Diagnostic Interview-Revised: a 2020.108840.
revised version of a diagnostic interview for caregivers of individuals with [16] M. Khosla, K. Jamison, A. Kuceyeski, M.R. Sabuncu, Ensemble learning with
possible pervasive developmental disorders, J. Autism Dev. Disord. 24 (5) (1994) 3D convolutional neural networks for functional connectome-based prediction,
659–685. NeuroImage 199 (2019) 651–662.
[8] S. Timimi, D. Milton, V. Bovell, S. Kapp, G. Russell, Deconstructing diagnosis: [17] Y. Liu, L. Xu, J. Li, J. Yu, X. Yu, Attentional connectivity-based prediction of
Four commentaries on a diagnostic tool to assess individuals for autism spectrum autism using heterogeneous rs-fMRI data from CC200 atlas, Exp. Neurobiol. 29
disorders, Autonomy (Birmingham, England) 1 (6) (2019). (1) (2020) 27–37, http://dx.doi.org/10.5607/en.2020.29.1.27.

9
X. Liu et al. Computers in Biology and Medicine 182 (2024) 109083

[18] N.C. Dvornek, P. Ventola, J.S. Duncan, Combining phenotypic and resting-state [35] Z. Khandan Khadem-Reza, M.A. Shahram, H. Zare, Altered resting-state func-
fMRI data for autism classification with recurrent neural networks, in: 2018 IEEE tional connectivity of the brain in children with autism spectrum disorder,
15th International Symposium on Biomedical Imaging, ISBI 2018, IEEE, 2018, Radiol. Phys. Technol. 16 (2) (2023) 284–291, http://dx.doi.org/10.1007/
pp. 725–728. s12194-023-00717-2.
[19] T. Eslami, V. Mirjalili, A. Fong, A.R. Laird, F. Saeed, ASD-DiagNet: a hybrid [36] F.E.-z. Bazay, A. Drissi El Maliani, Assessing the impact of preprocessing
learning approach for detection of autism spectrum disorder using fMRI data, pipelines on fMRI based autism spectrum disorder classification: ABIDE II results,
Front. Neuroinform. 13 (2019) 70, http://dx.doi.org/10.3389/fninf.2019.00070. in: International Conference on Engineering Applications of Neural Networks,
[20] F. Almuqhim, F. Saeed, ASD-SAENet: a sparse autoencoder, and deep-neural Springer, 2024, pp. 463–477, http://dx.doi.org/10.1007/978-3-031-62495-7_35.
network model for detecting autism spectrum disorder (ASD) using fMRI data, [37] L. Mahler, Q. Wang, J. Steiglechner, F. Birk, S. Heczko, K. Scheffler, G. Lohmann,
Front. Comput. Neurosci. 15 (2021) 654315, http://dx.doi.org/10.3389/fncom. Pretraining is all you need: A multi-atlas enhanced transformer framework for
2021.654315. autism spectrum disorder classification, in: International Workshop on Machine
[21] R. Anirudh, J.J. Thiagarajan, Bootstrapping graph convolutional neural networks Learning in Clinical Neuroimaging, Springer, 2023, pp. 123–132, http://dx.doi.
for autism spectrum disorder classification, in: ICASSP 2019-2019 IEEE Interna- org/10.1007/978-3-031-44858-4_12.
tional Conference on Acoustics, Speech and Signal Processing, ICASSP, IEEE, [38] R.C. Craddock, G.A. James, P.E. Holtzheimer III, X.P. Hu, H.S. Mayberg, A whole
2019, pp. 3197–3201, http://dx.doi.org/10.1109/ICASSP.2019.8683547. brain fMRI atlas generated via spatially constrained spectral clustering, Human
[22] Z. Sherkatghanad, M. Akhondzadeh, S. Salari, M. Zomorodi-Moghadam, M. Brain Mapp. 33 (8) (2012) 1914–1928, http://dx.doi.org/10.1002/hbm.21333.
Abdar, U.R. Acharya, R. Khosrowabadi, V. Salari, Automated detection of autism [39] R.M. Thomas, S. Gallo, L. Cerliani, P. Zhutovsky, A. El-Gazzar, G. van Wingen,
spectrum disorder using a convolutional neural network, Front. Neurosci. 13 Classifying autism spectrum disorder using the temporal statistics of resting-state
(2020) 1325, http://dx.doi.org/10.3389/fnins.2019.01325. functional MRI data with 3D convolutional neural networks, Front. Psychiatry
[23] B. Crosson, A. Ford, K.M. McGregor, M. Meinzer, S. Cheshkov, X. Li, D. Walker- 11 (2020).
Batson, R.W. Briggs, Functional imaging and related techniques: an introduction [40] F.Z. Subah, K. Deb, A comprehensive study on atlas-based classification of
for rehabilitation researchers, J. Rehabil. Res. Dev. 47 (2) (2010) vii. autism spectrum disorder using functional connectivity features from resting-
[24] M. Greicius, Resting-state functional connectivity in neuropsychiatric disorders, state functional magnetic resonance imaging, in: Neural Engineering Techniques
Curr. Opin. Neurol. 21 (4) (2008) 424–430. for Autism Spectrum Disorder, Elsevier, 2023, pp. 269–296, http://dx.doi.org/
[25] Y. Du, Z. Fu, V.D. Calhoun, Classification and prediction of brain disorders using 10.1016/B978-0-12-824421-0.00021-7.
functional connectivity: promising but challenging, Front. Neurosci. 12 (2018) [41] E. Yee, Identifying neural patterns and biomarkers of ASD through multi-phase
525. resting-state functional MRI analysis, in: 2023 11th International Conference
[26] A. Di Martino, D. O’connor, B. Chen, K. Alaerts, J.S. Anderson, M. Assaf, J.H. on Bioinformatics and Computational Biology, ICBCB, 2023, pp. 147–150, http:
Balsters, L. Baxter, A. Beggiato, S. Bernaerts, et al., Enhancing studies of the //dx.doi.org/10.1109/ICBCB57893.2023.10246705.
connectome in autism using the autism brain imaging data exchange II, Scient. [42] S.B. Eickhoff, K.E. Stephan, H. Mohlberg, C. Grefkes, G.R. Fink, K. Amunts,
Data 4 (1) (2017) 1–15, http://dx.doi.org/10.1038/sdata.2017.10. K. Zilles, A new SPM toolbox for combining probabilistic cytoarchitectonic
[27] Y. Kong, J. Gao, Y. Xu, Y. Pan, J. Wang, J. Liu, Classification of autism spectrum maps and functional imaging data, NeuroImage 25 (4) (2005) 1325–1335,
disorder by combining brain connectivity and deep neural network classifier, http://dx.doi.org/10.1016/j.neuroimage.2004.12.034.
Neurocomputing 324 (2019) 63–68. [43] A.R.T. Donders, G.J. Van Der Heijden, T. Stijnen, K.G. Moons, A gentle
[28] C.P. Chen, C.L. Keown, A. Jahedi, A. Nair, M.E. Pflieger, B.A. Bailey, R.-A. introduction to imputation of missing values, J. Clin. Epidemiol. 59 (10) (2006)
Müller, Diagnostic classification of intrinsic functional connectivity highlights 1087–1091.
somatosensory, default mode, and visual regions in autism, NeuroImage: Clin. 8 [44] M.J. Azur, E.A. Stuart, C. Frangakis, P.J. Leaf, Multiple imputation by chained
(2015) 238–245. equations: what is it and how does it work? Int. J. Methods Psychiatric Res. 20
[29] M. Plitt, K.A. Barnes, A. Martin, Functional connectivity classification of autism (1) (2011) 40–49, http://dx.doi.org/10.1002/mpr.329.
identifies highly predictive brain features but falls short of biomarker standards, [45] Y.-W. Chen, C.-J. Lin, Combining SVMs with various feature selection strate-
NeuroImage: Clin. 7 (2015) 359–366. gies, in: Feature extraction: foundations and applications, Springer, 2006, pp.
[30] M.A. Aghdam, A. Sharifi, M.M. Pedram, Diagnosis of autism spectrum disorders 315–324.
in young children based on resting-state functional magnetic resonance imaging [46] L. Rokach, Ensemble-based classifiers, Artif. Intell. Rev. 33 (2010) 1–39.
data using convolutional neural networks, J. Dig. Imag. 32 (2019) 899–918, [47] C. Craddock, Y. Benhajali, C. Chu, F. Chouinard, A. Evans, A. Jakab, B.S.
http://dx.doi.org/10.1007/s10278-019-00196-1. Khundrakpam, J.D. Lewis, Q. Li, M. Milham, et al., The neuro bureau preprocess-
[31] N. Payakachat, J.M. Tilford, W.J. Ungar, National Database for Autism Research ing initiative: open sharing of preprocessed neuroimaging data and derivatives,
(NDAR): big data opportunities for health services research and health technol- Front. Neuroinform. 7 (2013) 27.
ogy assessment, Pharmacoeconomics 34 (2) (2016) 127–138, http://dx.doi.org/ [48] A.V. Utevsky, D.V. Smith, S.A. Huettel, Precuneus is a functional core of the
10.1007/s40273-015-0331-6. default-mode network, J. Neurosci. 34 (3) (2014) 932–940.
[32] G. Li, M. Liu, Q. Sun, D. Shen, L. Wang, Early diagnosis of autism disease by [49] V.L. Cherkassky, R.K. Kana, T.A. Keller, M.A. Just, Functional connectivity in a
multi-channel CNNs, in: Machine Learning in Medical Imaging: 9th International baseline resting-state network in autism, NeuroReport 17 (16) (2006) 1687–1690,
Workshop, MLMI 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, http://dx.doi.org/10.1097/01.wnr.0000239956.45448.4c.
September 16, 2018, Proceedings 9, Springer, 2018, pp. 303–309, http://dx.doi. [50] T. Watanabe, N. Yahata, O. Abe, H. Kuwabara, H. Inoue, Y. Takano, N. Iwashiro,
org/10.1007/978-3-030-00919-9_35. T. Natsubori, Y. Aoki, H. Takao, et al., Diminished medial prefrontal activity
[33] M. Khodatars, A. Shoeibi, D. Sadeghi, N. Ghaasemi, M. Jafari, P. Moridian, behind autistic social judgments of incongruent information, PLoS One 7 (6)
A. Khadem, R. Alizadehsani, A. Zare, Y. Kong, A. Khosravi, S. Nahavandi, S. (2012) e39561.
Hussain, U.R. Acharya, M. Berk, Deep learning for neuroimaging-based diagnosis [51] V. Menon, Developmental pathways to functional brain networks: emerging
and rehabilitation of Autism Spectrum Disorder: A review, Comput. Biol. Med. principles, Trends in Cognitive Sciences 17 (12) (2013) 627–640, http://dx.doi.
139 (2021) 104949, http://dx.doi.org/10.1016/j.compbiomed.2021.104949. org/10.1016/j.tics.2013.09.015.
[34] M.P. Van Den Heuvel, H.E.H. Pol, Exploring the brain network: a review on [52] H. Koshino, P.A. Carpenter, N.J. Minshew, V.L. Cherkassky, T.A. Keller,
resting-state fMRI functional connectivity, Eur. Neuropsychopharmacol. 20 (8) M.A. Just, Functional connectivity in an fMRI working memory task in
(2010) 519–534. high-functioning autism, Neuroimage 24 (3) (2005) 810–821.
[53] P. Domingos, A few useful things to know about machine learning, Commun.
ACM 55 (10) (2012) 78–87, http://dx.doi.org/10.1145/2347736.2347755.

10