[Paper Review] Deep Learning for Neuroimaging-based Diagnosis and Rehabilitation of Autism Spectrum Disorder: A Review

Marjane Khodatars, Afshin Shoeibi+13 more|arXiv (Cornell University)|Jul 2, 2020

Autism Spectrum Disorder Research309 references267 citations

TL;DR

This review synthesizes deep learning (DL) applications in neuroimaging-based diagnosis and rehabilitation of Autism Spectrum Disorder (ASD), evaluating DL models across structural (sMRI, DTI) and functional (fMRI, EEG, fNIRS) neuroimaging modalities. It identifies key DL architectures—such as 3D-CNNs, LSTMs, autoencoders, and GCNs—that achieve high classification accuracy (up to 95%) in distinguishing ASD from neurotypical controls, while also assessing DL-driven rehabilitation tools and outlining critical challenges in data, interpretability, and clinical translation.

ABSTRACT

Accurate diagnosis of Autism Spectrum Disorder (ASD) followed by effective rehabilitation is essential for the management of this disorder. Artificial intelligence (AI) techniques can aid physicians to apply automatic diagnosis and rehabilitation procedures. AI techniques comprise traditional machine learning (ML) approaches and deep learning (DL) techniques. Conventional ML methods employ various feature extraction and classification techniques, but in DL, the process of feature extraction and classification is accomplished intelligently and integrally. DL methods for diagnosis of ASD have been focused on neuroimaging-based approaches. Neuroimaging techniques are non-invasive disease markers potentially useful for ASD diagnosis. Structural and functional neuroimaging techniques provide physicians substantial information about the structure (anatomy and structural connectivity) and function (activity and functional connectivity) of the brain. Due to the intricate structure and function of the brain, proposing optimum procedures for ASD diagnosis with neuroimaging data without exploiting powerful AI techniques like DL may be challenging. In this paper, studies conducted with the aid of DL networks to distinguish ASD are investigated. Rehabilitation tools provided for supporting ASD patients utilizing DL networks are also assessed. Finally, we will present important challenges in the automated detection and rehabilitation of ASD and propose some future works.

Motivation & Objective

• To systematically review deep learning (DL) techniques applied to neuroimaging data for automated diagnosis of Autism Spectrum Disorder (ASD).
• To evaluate the performance and architecture of DL models in classifying ASD using structural and functional neuroimaging (e.g., MRI, fMRI, EEG, fNIRS).
• To assess DL-based rehabilitation tools designed to support ASD patients through personalized therapy and intervention systems.
• To identify key challenges in data scarcity, model interpretability, generalizability, and clinical integration of DL in ASD applications.
• To propose future research directions for improving the reliability, scalability, and real-world deployment of DL in ASD diagnosis and therapy.

Proposed method

• Conducted a systematic literature review using databases (IEEE Xplore, ScienceDirect, SpringerLink, ACM) with keywords: 'ASD', 'Autism Spectrum Disorder', and 'Deep Learning' up to June 2020.
• Categorized and analyzed 128 studies applying DL to ASD diagnosis and rehabilitation, focusing on neuroimaging modalities including sMRI, DTI, fMRI, EEG, fNIRS, MEG, and ECoG.
• Evaluated DL architectures such as 3D-CNNs, 2D-CNNs, LSTMs, autoencoders (AE, VAE, SAE), GCNs, and hybrid models (e.g., CNN-LSTM, 3D-UNet) for classification and feature learning.
• Reviewed loss functions (BCE, CE, MSE, NLLF, proposed losses), optimizers (Adam, SGD, Adadelta), activation functions (ReLU, ELU, Sigmoid, Softmax), and regularization (dropout, batch normalization).
• Assessed model performance using metrics like accuracy, sensitivity, specificity, AUC, and F1-score, with reported values up to 95% accuracy in some studies.
• Evaluated rehabilitation systems using DL for personalized therapy, including brain-computer interfaces, emotion recognition, and behavioral intervention planning.

Experimental results

Research questions

• RQ1 Which deep learning architectures demonstrate the highest diagnostic accuracy in distinguishing ASD from neurotypical controls using neuroimaging data?
• RQ2 How do different neuroimaging modalities (e.g., fMRI, EEG, sMRI) compare in their suitability for DL-based ASD classification?
• RQ3 What are the most effective DL-based approaches for developing personalized rehabilitation tools for ASD patients?
• RQ4 What are the major technical and clinical challenges hindering the deployment of DL in real-world ASD diagnosis and therapy?
• RQ5 How can interpretability, generalizability, and data efficiency be improved in DL models for ASD applications?

Key findings

• 3D-CNNs and hybrid models (e.g., 3D-CNN + LSTM) achieved high diagnostic accuracy, with one study reporting up to 95% accuracy in classifying ASD using fMRI data.
• Autoencoders (AE), stacked autoencoders (SAE), and variational autoencoders (VAE) demonstrated strong feature extraction capabilities, particularly in dimensionality reduction and representation learning.
• LSTM and GRU-based networks showed strong performance in modeling temporal dynamics in EEG and fMRI time-series data, with accuracy rates exceeding 90% in several studies.
• Pretrained models like ResNet-18 and VGG-based networks (e.g., VGGNets + ASDNet) achieved competitive performance (AUC > 0.90) in ASD classification tasks.
• Hybrid architectures such as 3D-UNet and FastSurfer CNN showed high segmentation and classification accuracy (AUC > 0.92) in structural MRI analysis.
• Despite high performance, challenges remain in model interpretability, data scarcity, and generalization across diverse populations, with most studies relying on small, single-center datasets.