PaperlayerView exact revision on arXiv

Source-linked AI summary

Automatic autism spectrum disorder detection using artificial intelligence methods with MRI neuroimaging: A review

Parisa Moridian, Navid Ghassemi, Mahboobeh Jafari, Salam Salloum-Asfar, Delaram Sadeghi, Marjane Khodatars, Afshin Shoeibi, Abbas Khosravi, Sai Ho Ling, Abdulhamit Subasi, Roohallah Alizadehsani, Juan M. Gorriz, Sara A Abdulla, U. Rajendra Acharya

arXiv:2206.11233v3q-bio.NCcs.LGeess.IV

TL;DR

MRI-based ASD diagnosis uses functional and structural imaging, but interpreting scans can be demanding for clinicians. This review surveys AI-based diagnostic systems and compares reported patterns across ML and DL studies. It describes DL use as growing and notes limited access to large public datasets spanning different ASD types.

  • Problem

    MRI interpretation can be time-consuming for clinicians, while the review notes limited access to large public datasets representing different ASD types.

    Source: §7 · p. 25, §8 · p. 26

  • Method

    The paper reviews AI-based computer-aided systems for MRI-based ASD detection, covering ML studies and comparing them with DL studies.

    Source: §1 · p. 3, §8 · p. 27

  • Results

    DL architectures are reported as more used than ML models in studies of automated ASD diagnosis with MRI, and rs-fMRI is the most-used modality across both ML and DL studies.

    Source: §6 · pp. 22, 24

  • Takeaways & Limitations

    The review presents AI methods and MRI neuroimaging as promising for developing applied ASD-detection software and supporting quicker early diagnosis.

    Source: §8 · p. 27

    Access to large public MRI datasets containing different types of ASD remains a challenge identified by the authors.

    Source: §7 · p. 25

Abstract

from arXiv · show

Autism spectrum disorder (ASD) is a brain condition characterized by diverse signs and symptoms that appear in early childhood. ASD is also associated with communication deficits and repetitive behavior in affected individuals. Various ASD detection methods have been developed, including neuroimaging modalities and psychological tests. Among these methods, magnetic resonance imaging (MRI) imaging modalities are of paramount importance to physicians. Clinicians rely on MRI modalities to diagnose ASD accurately. The MRI modalities are non-invasive methods that include functional (fMRI) and structural (sMRI) neuroimaging methods. However, diagnosing ASD with fMRI and sMRI for specialists is often laborious and time-consuming; therefore, several computer-aided design systems (CADS) based on artificial intelligence (AI) have been developed to assist specialist physicians. Conventional machine learning (ML) and deep learning (DL) are the most popular schemes of AI used for diagnosing ASD. This study aims to review the automated detection of ASD using AI. We review several CADS that have been developed using ML techniques for the automated diagnosis of ASD using MRI modalities. There has been very limited work on the use of DL techniques to develop automated diagnostic models for ASD. A summary of the studies developed using DL is provided in the Supplementary Appendix. Then, the challenges encountered during the automated diagnosis of ASD using MRI and AI techniques are described in detail. Additionally, a graphical comparison of studies using ML and DL to diagnose ASD automatically is discussed. We suggest future approaches to detecting ASDs using AI techniques and MRI neuroimaging.

1. Introduction

ASD manifests in childhood and can affect communication, cognition, and social behavior; MRI offers structural and functional information, but interpreting scans can burden clinicians. The review frames AI-based computer-aided systems as a way to support MRI-based ASD diagnosis.

1. Introduction

  • ASD manifests in childhood and is associated with challenges in communication, cognitive skills, and social behavior; early diagnosis and treatment are presented as important to quality of life.The passage reports prevalence of one in 44 children aged 8 years and about four times greater prevalence among boys than girls.Source: §1 · p. 2
  • fMRI non-invasively records brain activity and supports functional-connectivity analysis, while sMRI examines anatomy and is described as cost-effective and widely available.The review also notes DTI for examining interactions between brain regions.Source: §1 · p. 2
  • MRI artifacts, multiple slices, protocol differences, and interpretation demands can make accurate clinical diagnosis time-consuming and difficult, especially for inexperienced physicians.Physician fatigue may also contribute to incorrect diagnoses.Source: §1 · p. 3
  • AI-based computer-aided systems are reviewed as support for MRI-based ASD detection, with coverage of ML studies and a comparison to DL studies summarized in the supplementary appendix.The paper notes that ML feature extraction, dimension reduction, and classification choices can require trial and error, while DL uses deep layers for feature extraction and classification.Source: §1 · p. 3

2. Search Strategy Based on PRISMA Guideline

The review searched for ASD diagnosis papers using MRI and AI, following PRISMA selection procedures. It included papers published from 2016 to 2022 and retained 233 papers after screening.

2. Search Strategy Based on PRISMA Guideline

  • The search covered papers published from 2016 to 2022 and used databases including IEEE, Wiley, Frontiers, ScienceDirect, SpringerLink, ACM, and ArXiv, plus Google Scholar.Search terms included ASD classification, feature extraction, fMRI, sMRI, and deep-learning-related terms.Source: §2 · pp. 3–4
  • 233 papers were retained after three screening stages reduced the 316 downloaded papers by excluding out-of-scope papers, papers without MRI datasets, and papers without AI techniques.The exclusions were 34, 28, and 21 papers, respectively.Source: §2 · p. 4
  • Paper selection followed the PRISMA protocol, with inclusion and exclusion criteria presented in Table 1 and the selection process illustrated in Figure 1.The paper describes selection at three levels.Source: §2 · p. 4, Figure 1 · p. 4

3. Artificial Intelligence Techniques for ASD diagnosis

This section situates the review among prior work on AI-based ASD diagnosis and identifies its focus on MRI studies using both ML and DL. It also presents challenges and future directions for this research area.

3. Artificial Intelligence Techniques for ASD diagnosis

  • Prior work reviewed brain-structure findings in sMRI, ML methods across sMRI and fMRI, and feature extraction, selection, reduction, training, testing, and evaluation approaches.Pagnozzi et al. reported structural differences; Nogay et al. reported higher diagnostic accuracy at younger ages; Xu et al. outlined ML methods and evaluation parameters.Source: §3 · p. 5
  • Prior reviews covered sMRI, fMRI, combined MRI, unsupervised ML, and DL approaches, while this review focuses on ASD diagnosis using MRI data and ML and DL methods.The cited reviews include Pagnozzi et al., Nogay et al., Xu et al., Parlett-Pellerit et al., Rahman et al., Eslami et al., and Khodatars et al..Source: §3 · pp. 5–6
  • The review reports coverage of ASD diagnosis articles from 2010 to 2022 and discusses challenges and future work for diagnosis from MRI modalities.The authors state that, to their knowledge, no similar review had been provided.Source: §3 · p. 6

4. CADS for ADS diagnosis by MRI Neuroimaging Modalities

This section surveys MRI-based computer-aided diagnosis systems, covering machine-learning pipelines and methods alongside deep-learning systems. It summarizes imaging features, classification approaches, and diagnostic performance metrics reported across studies.

4. CADS for ADS diagnosis by MRI Neuroimaging Modalities

  • Machine-learning CADS process MRI data through preprocessing, feature extraction, feature reduction or selection, and classification; deep-learning systems perform feature extraction through deep layers without user intervention.The section says that with more input data, deep-learning CADS may enable development of actual ASD-detection software.Source: §4 · p. 6

4.1. MRI Neuroimaging ASD Datasets

  • ABIDE is described as the most complete freely available MRI database for automatic ASD diagnosis, with ABIDE I and II containing 1,112 and 1,114 datasets, respectively.The dataset includes sMRI, resting-state fMRI, and phenotypic data; ABIDE I also offers preprocessed data through the Preprocessed Connectomes Project (PCP).Source: §4.1 · p. 6

4.2. Preprocessing Techniques for fMRI and sMRI Modalities

  • Standard preprocessing is necessary for fMRI and sMRI CADS; fMRI steps include motion correction and atlas registration, while sMRI steps include denoising and segmentation.The passage warns that improper fMRI preprocessing reduces automated diagnostic performance and names BET, FSL, SPM, and FreeSurfer as commonly used software packages.Source: §4.2 · pp. 7–8

4.2.1. Pipeline Methods

  • ABIDE pipelines provide data preprocessed with standard procedures to help avoid output variations between studies; popular options include NIAK, DPARSF, CPAC, and CCS.Source: §4.2.1 · p. 8

4.3. Feature Extraction

  • MRI feature extraction schemes include statistical, texture, morphological, nonlinear, graph, functional-connectivity, and structural-connectivity features; connectivity methods are most frequently used in summarized studies.Statistical features are described as informative and quick to extract, but unable to reveal nonlinear or subtle data patterns.Source: §4.3 · pp. 9–10

4.4. Feature Reduction / Selection Methods

  • Feature reduction and selection address the curse of dimensionality; reviewed methods include PCA, RFE, t-test, autoencoder, conditional random forest, chi-squared, and LASSO.PCA represents data using principal components that preserve variance, while RFE iteratively eliminates less-important features.Source: §4.4 · p. 10

4.5. Classification Methods

  • Classification is the final machine-learning CADS stage, and SVM, LDA, KNN, and random forest are among the most popular classifiers described for ASD detection.SVM separates points with a hyperplane, random forests combine tree votes, KNN assigns labels from nearby points, and LDA partitions data into class regions.Source: §4.5 · pp. 11–12
  • The catalog includes functional-connectivity graph feature selection with an MMoE model, Bag-of-Feature extraction, and functional-connectivity low-rank representation decomposition.Source: §4.5 · p. 17
  • Other entries report Acc=76.3, Sen=79.2, Spec=63.9; Acc=81, Sen=81, Spec=86; and Acc=73.44, Sen=75.79, Spec=69.52.Source: §4.5 · p. 17
  • For study on ABIDE data (250 ASD, 218 HC), using rs-fMRI with AAL Atlas + CPAC, Acc=90.85, Sen=90.86, and Spec=90.90 are reported.Source: §4.5 · p. 17
  • Additional reported metrics include Acc=74.86, Prec=76.07, Recall=71.67, and Acc=65.03.Source: §4.5 · p. 18
  • [299] reported specificity=86 and sensitivity=88 using RFE-SVM and SVM classification.The study used GLM-derived features.Source: §4.5 · p. 18
  • Accuracy exceeded 96.3 in [304], which used fMRI functional ROIs and connectivity features with logistic regression.The listed features included functional connectivity and seed-based connectivity.Source: §4.5 · p. 18
  • [317] reported specificity >34 and sensitivity >92 for MRI-based classification.Source: §4.5 · p. 19
  • [320] reported AUC=82, while reported AUC=81 for sMRI-based SVM classifiers.[320] used gray-matter volumes with RFE; used regional features.Source: §4.5 · p. 19
  • [323] reported AUC=74 using sMRI regional features with an SVM.Source: §4.5 · p. 19

5. Challenges in detecting ASD with MRI neuroimaging modalities and AI techniques

The review identifies dataset limitations, unavailable multimodal ASD datasets, challenges in AI algorithms, and inadequate hardware resources as barriers to automated ASD detection.

Dataset limitations

  • Available MRI datasets classify only ASD and control, limiting research on systems that distinguish different ASD types.Datasets covering different ASD types could support accurate diagnosis of those types.Source: §5 · p. 21
  • ASD studies lack multimodal neuroimaging datasets, constraining research on modality fusion despite its reported diagnostic benefit in brain disorders.The review names fMRI-MEG, MRI-PET, and EEG-fMRI as important multimodalities and says such datasets might lead to practical ASD studies.Source: §5 · p. 21

AI algorithm challenges

  • ML-based CADS are time-consuming and complex to design, while DL automates feature extraction through classification but may perform worse because features lack supervision.The review says combining ML and DL methods may yield promising ASD diagnosis results.Source: §5 · p. 21

Hardware challenges

  • Limited access to suitable hardware makes storing and processing large public datasets such as ABIDE challenging on ordinary computers.Cloud-server use for CAD implementation has not been seriously investigated; compact CNN models require fewer hardware resources than other CNN methods.Source: §5 · p. 21

6. Discussion

The review compares MRI-based ASD detection studies using ML and DL across publication activity, datasets, modalities, atlases, preprocessing pipelines, and classifiers. It reports greater recent DL activity and identifies the most-used choices in several of these areas.

Comparison between the numbers of papers published each year for ML and DL research

  • DL models are used in more recent MRI-based ASD detection studies than ML models, and research interest in DL has grown significantly.The reviewed studies began in 2017; the authors describe DL-based CADS as promising for future software development.Source: §6 · p. 22, Figure 4 · p. 22

Comparison between the numbers of datasets used in the ML and DL research

  • ABIDE datasets are used more often than other datasets in automated ASD diagnosis studies because they offer many subjects and different MRI modalities.The review identifies ABIDE I and ABIDE II as the two available ABIDE datasets.Source: §6 · pp. 22–23, Figure 5 · p. 23

Comparison between the numbers of neuroimaging modalities used in the ML and DL research

Comparison between the numbers of Atlases used in the ML and DL research

Comparison between the numbers of pipelines used in the ML and DL researches

  • CPAC is the most widely used pipeline among the reviewed ASD detection studies.The review describes pipelines as part of MRI preprocessing.Source: §6 · p. 25, Figure 8 · p. 25

Comparison between the numbers of classification methods in the ML and DL research

  • Softmax is the most-used classification method in DL architectures, while SVM is the most widely applied method in ML studies.Classification is described as the final CADS step.Source: §6 · p. 25, Figure 9 · p. 25

7. Future Works

The future-work agenda calls for broader ASD and MRI datasets, new ML and DL methods, and feature fusion for automated diagnosis. It also identifies multimodal data, including fMRI-EEG, as a direction for further study.

7. Future Works

  • Public MRI datasets covering different ASD types are needed; ABIDE has relatively few DTI cases despite DTI’s use in ASD detection.The review says broader datasets could support more studies and comparisons across researchers.Source: §7 · pp. 25–26
  • Multimodal datasets such as fMRI-EEG are proposed to support studies of ASD diagnosis using different AI methods.The cited clinical studies indicate a pivotal role for multimodal neuroimaging in ASD diagnosis.Source: §7 · p. 26
  • For ML, proposed directions include fuzzy-based feature extraction and classification, fuzzy graph models, connectivity-based features, and graph-theory classification.The review notes that fuzzy-based feature extraction has not been used for ASD diagnosis in the studies it summarizes.Source: §7 · p. 26
  • For DL, the review proposes exploring graph theory, representation learning, zero-shot learning, Q-learning, attention learning, and advanced adversarial-network models.The reviewed studies used standard DL methods.Source: §7 · p. 26
  • Feature fusion could combine DL features extracted from MRI images with ML features for automated ASD detection.The review presents this as a route to high performance, not as an established result.Source: §7 · p. 26

8. Conclusion

The conclusion frames MRI-based AI as a response to the time and difficulty of ASD diagnosis and reviews ML and DL research across MRI modalities. It also identifies research challenges and future directions, while comparing publication, dataset, modality, preprocessing, atlas, and classifier patterns across studies.

8. Conclusion

  • MRI-based ASD diagnosis can be time-consuming and challenging, and tiredness or MRI noise may contribute to clinicians’ incorrect diagnoses.Source: §8 · p. 26
  • The review covers AI-based ASD detection across MRI modalities, including ML CADS, prior DL-review work, and challenges such as dataset access, multimodal data, algorithms, and hardware.It positions this work alongside earlier reviews of DL and ML using different neuroimaging modalities.Source: §8 · p. 27
  • The discussion compares annual publication counts, datasets, MRI modalities, atlases, preprocessing pipelines, and classifier algorithms in ML and DL studies.Source: §8 · p. 27
  • Future directions include MRI datasets and AI methods, with feature fusion also proposed; the review describes the methods as promising for real ASD-detection software and quicker early diagnosis.Source: §8 · p. 27

Appendix A:

Appendix A catalogs deep-learning studies of automated ASD diagnosis using MRI, summarizing their datasets, imaging modalities, and model approaches.

Appendix A:

  • The table covers task and resting-state fMRI as well as sMRI, using clinical, HCP, ABIDE, and NDAR datasets [197],, [203], [234], [237].The entries also include ABIDE-I and ABIDE-II data and multiple atlas or processing-pipeline choices.Source: Table 3 · p. 28, Appendix A · pp. 28–29
  • The listed methods include G-CNNs, BrainNet, DAE and SAE, a DANN with MLP, self-attention, fusion and aggregation, and CNN variants including 1D-CNNs [207], [214], [219].Source: Appendix A · pp. 28–29
  • The catalog also includes studies combining imaging modalities, including fMRI and sMRI, and entries using clinical imaging data [233],,.Other listed work uses task fMRI or all imaging modalities,.Source: Appendix A · pp. 29, 31

Appendix A

  • The listed studies include DBN and CNN-based approaches using ABIDE data, with rs-fMRI or sMRI/fMRI and atlas or preprocessing pipelines [291], [293].The entries specify a three-RBM-layer DBN, a 2D-CNN, and CNNGLasso.Source: Appendix A · p. 31
  • Graph-based models include p-GCN, s-GCN, ss-GCN, MGCN, and LSTM-ANN architectures across ABIDE II, KKI, and clinical rs-fMRI/DTI data [294],, [296].The LSTM-ANN entry describes two hidden layers for both P-ANN and A-ANN.Source: Appendix A · pp. 31–32
  • Other entries list 1D- and 3D-CNNs, a multi-atlas graph convolutional network, and an MHATC attention encoder for MRI studies, [329], [330], [331].The MHATC entry combines multi-head attention, temporal unsupervised pretraining, and MLP supervised fine-tuning.Source: Appendix A · p. 32
  • The appendix also lists deep-fusion and stacked-autoencoder approaches, CAM visualization architectures, and CNN and SSAE-plus-MLP models [332], [333], [334], [341],.The CAM entry includes 2D CAM, 3D CAM, and 3D Grad-CAM.Source: Appendix A · p. 32
  • Further graph designs include invertible dynamic GCN, functional graph discrimination, and DFS-aggregation graph convolutional networks [335], [336], [337].The entries describe invertible blocks, graph construction and convolution layers, and combined DFS and aggregation.Source: Appendix A · p. 32
  • Later entries describe multi-scale graph representation and feature-fusion methods, alongside multi-view graph learning with view-consistency regularization [339], [340], [343].The listed components include multi-scale functional-connectivity networks and embedding learning for different brain-network views.Source: Appendix A · p. 32

Related papers

Loading 2206.11233v3…