Source-linked AI summary
Deep Learning for Neuroimaging-based Diagnosis and Rehabilitation of Autism Spectrum Disorder: A Review
Marjane Khodatars, Afshin Shoeibi, Delaram Sadeghi, Navid Ghassemi, Mahboobeh Jafari, Parisa Moridian, Ali Khadem, Roohallah Alizadehsani, Assef Zare, Yinan Kong, Abbas Khosravi, Saeid Nahavandi, Sadiq Hussain, U. Rajendra Acharya, Michael Berk
TL;DR
ASD diagnosis lacks a pathophysiological marker, making accurate assessment and effective rehabilitation important. This review synthesizes DL-based neuroimaging diagnosis and rehabilitation studies, finding CNNs especially prominent and identifying dataset, generalization, and hardware challenges.
Problem
ASD diagnosis lacks a pathophysiological marker and relies on psychological criteria, motivating investigation of automated diagnosis and rehabilitation approaches.
Method
The paper reviews DL networks, neuroimaging-based ASD computer-aided diagnosis systems, and DL-based rehabilitation tools.
Results
Acc=95.7 was reported for the listed ASD detection study using a 1D CNN-LSTM with bagging.
Takeaways & Limitations
The review identifies CNN as the most popular architecture, with autoencoders and RNNs also yielding favorable results, and proposes wearable low-power rehabilitation devices as future tools.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
ASD diagnosis is difficult because it lacks a pathophysiological marker and relies on psychological criteria, while neuroimaging offers structural and functional information that DL can analyze. This review surveys DL-based ASD diagnosis and rehabilitation, using literature identified through June 3, 2020.
- I. INTRODUCTION: ASD diagnosis is difficult because no pathophysiological marker exists, so clinicians rely on psychological criteria and behavioral evaluations.Behavioral tools include clinical observations, medical history, diagnostic instructions, and growth and intelligence tests.
- I. INTRODUCTION: Structural neuroimaging studies brain anatomy and structural connections, whereas functional neuroimaging examines brain activity and functional connections.MRI and DTI-MRI support structural analysis; EEG and fMRI support functional investigation.
- I. INTRODUCTION: The review covers DL networks, neuroimaging-based CAD systems, rehabilitation tools, challenges, and future work for ASD diagnosis and rehabilitation.The literature search used IEEE Xplore, ScienceDirect, SpringerLink, ACM, and other venues with ASD and Deep Learning keywords through June 3, 2020.
- I. INTRODUCTION: DL-based CAD systems integrate feature extraction and classification within the model, unlike traditional AI-based CAD systems that use separate stages.DL networks require large datasets to learn intricate patterns.
- I. INTRODUCTION: ABIDE is described as the most complete free dataset, containing sMRI, rs-fMRI, and phenotypic data across 17 international sites.ABIDE-I includes 1112 datasets: 539 individuals with ASD and 573 healthy individuals.
1) Standard (Low-level) fMRI preprocessing steps:
Low-level fMRI preprocessing applies standardized corrections and normalization before analysis, while atlas registration extracts regional representations. Pipelines and higher-level techniques support comparable and potentially more accurate DL-based ASD recognition.
- 1) Standard (Low-level) fMRI preprocessing steps:: Standard fMRI preprocessing includes brain extraction, spatial smoothing, temporal filtering, motion correction, slice timing correction, intensity normalization, and atlas registration.FSL, BET, FreeSurfer, and SPM are among the toolboxes used for these steps.
- 1) Standard (Low-level) fMRI preprocessing steps:: Brain extraction removes the skull and cerebellum while preserving brain tissue.
- 1) Standard (Low-level) fMRI preprocessing steps:: Spatial smoothing averages neighboring voxel signals because nearby voxels are usually related in function and blood supply.
- 1) Standard (Low-level) fMRI preprocessing steps:: Temporal filtering removes unwanted voxel time-series components without impairing the signal of interest.
- 1) Standard (Low-level) fMRI preprocessing steps:: Motion and slice-timing correction align image coordinates and establish a common reference time across voxels.
- 1) Standard (Low-level) fMRI preprocessing steps:: Intensity normalization compensates for global deviations within and between recording sessions, while atlas registration partitions images into ROIs for mean time-series extraction.
- 1) Standard (Low-level) fMRI preprocessing steps:: NIAK, DPARSF, CPAC, and CCS provide ABIDE preprocessing pipelines with broadly analogous steps, while sliding windows, augmentation, connectivity matrices, and FFT are higher-level techniques.Higher-level preprocessing methods are applied after standard preprocessing and may enhance ASD recognition accuracy.
1) Convolutional Neural Networks (CNNs) :
The review describes CNN, DBN, AE, RNN, and hybrid architectures for ASD detection, emphasizing how their structures address spatial, temporal, or unsupervised feature-learning needs. Network choice also involves data requirements and parameter efficiency.
- 1) Convolutional Neural Networks (CNNs) :: CNNs extract spatially dependent features with convolutional filters while reducing parameters, supporting training on smaller databases.The review distinguishes 1D-, 2D-, and 3D-CNNs according to input structure.
- 1) Convolutional Neural Networks (CNNs) :: 3D-CNNs require much larger training datasets and generally offer less benefit than 1D-CNNs and 2D-CNNs in the reviewed context.Pre-training extensively used in 2D networks cannot be used for 3D-CNNs according to the passage.
- 2) Deep Belief Networks (DBNs):: DBNs learn features without supervision through greedy layer-wise training, then provide features or initial weights for classification.
- 3) Autoencoders (AEs):: Autoencoders encode inputs into lower-dimensional latent representations and decode them to reconstruct the original data.The coder can serve as an unsupervised feature extractor, with sparse and denoising variants used to limit memorization.
- 4) Recurrent Neural Networks (RNNs):: RNNs address distant and variable-length temporal dependencies, with LSTM structures extracting long-term and short-term dependencies.GRU and GRU-D are described as related structures for improving sequence handling and finding lost data.
- CNN-RNN: CNN-RNN networks combine convolutional layers for spatial dependencies with RNNs for temporal dependencies, benefiting multidimensional time series such as video.
- CNN-AE: CNN-AE combinations reduce parameters, support higher-dimensional data, and extract more information without changing database size.Similar structures are widely used for image segmentation, with unsupervised variants also applicable.
V. DEEP LEARNING TECHNIQUES FOR ASD REHABILITATION
DL-based rehabilitation tools support ASD patients through mobile and computer applications, robotic devices, cloud systems, and eye tracking. The review presents these tools as a varied set of assistive approaches.
- V. DEEP LEARNING TECHNIQUES FOR ASD REHABILITATION: DL-based rehabilitation tools for ASD include mobile applications, computer applications, robotic devices, cloud systems, and eye-tracking systems.These tools are intended to help patients recover after treatment and are discussed across multiple rehabilitation settings.
A. Mobile and Software Applications
Mobile software applications support ASD rehabilitation by using facial images to assist children with social-emotional understanding. The reviewed iOS design photographs faces under varied angles and brightness conditions.
- Mobile and Software Applications: A smart iOS application was designed to assist children with ASD using facial images.The application design is represented in Figure 11.
- Mobile and Software Applications: Faces are photographed at different angles and brightness levels for the application.
- Mobile and Software Applications: A similar study achieved 78.32% accuracy.
B. Cloud Systems
A cloud-oriented rehabilitation approach combines convolutional and recurrent deep-learning models with inertial sensing to identify stereotypical motor movements. CNNs encode multi-sensor time series, while LSTMs capture temporal patterns before ensemble voting.
- Cloud Systems: A DL application identifies stereotypical motor movements using multi-axis inertial measurement units.
- Cloud Systems: A CNN transforms multi-sensor time series into feature space before an LSTM extracts temporal patterns.
- Cloud Systems: Classifier-selection voting combines an ensemble of the best base learners.
C. Eye Tracking
The review covers deep-learning approaches for ASD-related eye tracking and neuroimaging, including saliency prediction, MRI-based classification, and rehabilitation tools. Reported systems use varied datasets, preprocessing pipelines, architectures, and performance measures.
- Eye Tracking: A saliency-prediction model uses a DCN, single-side clipping, and an eye-tracking dataset for autistic children.The fixation density map is used with the saliency map to optimize the loss function.
- Eye Tracking: 96.88% accuracy was reported for an SNCAE-based system using s-MRI shape features.
- Eye Tracking: 79.2% accuracy was reported for normalized features evaluated with an autoencoder and DNN.
- Eye Tracking: The review also catalogs rehabilitation tools and the number of DL tools used across diagnosis and rehabilitation studies.Figure 13 summarizes tool usage, while Table II covers rehabilitation papers using DL algorithms.
VI. DISCUSSION
The discussion finds CNNs to be the most popular DL architecture in reviewed ASD detection studies, with autoencoders and RNNs also producing favorable results. Keras is widely used, while Softmax is a common classifier because it is differentiable and computationally inexpensive.
- VI. DISCUSSION: Keras is used in the majority of studies because its high-level API simplifies model construction and supports powerful backends.Pre-trained models and publicly available code also contribute to its popularity.
- VI. DISCUSSION: CNN is the most popular architecture in reviewed ASD detection studies and has achieved more promising results than other deep methodologies.
- VI. DISCUSSION: Autoencoders and RNNs have also yielded favorable results in the reviewed studies.
- VI. DISCUSSION: Softmax is among the most widely used classification algorithms because it is differentiable across its domain and computationally less expensive.
VII. CHALLENGES
The review identifies database and algorithmic barriers to deep-learning-based ASD diagnosis and rehabilitation. Public datasets largely cover only ASD-versus-healthy classification, while multimodal data and real-world hardware remain limited.
- Database problems: Public brain structural and functional datasets generally support only two-class ASD-versus-healthy studies, limiting investigation of ASD subtypes.The review states that researchers cannot broaden investigations to all ASD subtypes because public datasets are restricted to two classes.
- Database problems: Freely available EEG and fNIRS datasets are scarce, despite these modalities being among the cheapest and most pragmatic functional neuro-screening options.
- Database problems: Multimodality databases such as EEG-fMRI are unavailable for evaluating whether combined imaging information improves ASD detection.Although ABIDE contains fMRI and sMRI data, the review says their combination for DL-based ASD diagnosis had not yet been investigated.
- Algorithmic and implementation problems: Real-world deployment of DL rehabilitation systems remains constrained by hardware-resource requirements despite tools such as Google Colab improving processing power.
VIII. CONCLUSION AND FUTURE WORKS
The review concludes that deep learning has been applied to ASD diagnosis and rehabilitation through several network families and proposes wearable, low-power tools as a future direction. It highlights accuracy, generalization, adaptability, data scarcity, and hardware demands as continuing concerns.
- Conclusion: Deep CNNs, RNNs, autoencoders, and CNN-RNN networks have been used to improve ASD diagnosis and rehabilitation systems.
- Future works: Future ASD systems should improve accuracy, generalization to differing data, adaptation to real-world challenges, and hardware efficiency.
- Future works: Deep reinforcement learning and GANs are proposed as future approaches for improving CADS performance and addressing medical-data scarcity.
- Future works: Wearable, reliable, accurate, low-power devices such as camera-equipped smart glasses are proposed to help children with ASD communicate with their surroundings.The glasses would capture environmental images and use DL to produce meaningful images for the children.
APPENDIX A STATISTICAL METRICS
The appendix defines statistical measures used to evaluate ASD classification systems, including accuracy, TPR, ROC curves, and AUC. These metrics describe classification correctness and performance across thresholds.
- Metric definitions: Evaluation metrics use TP, TN, FP, and FN to distinguish correct and incorrect positive and negative classifications.
- Metric definitions: Accuracy measures the proportion of predictions classified correctly among all positive and negative outcomes.
- ROC analysis: The ROC curve plots true positive rate against false positive rate across all classification thresholds.
- ROC analysis: AUC summarizes performance across all possible classification thresholds and ranges from 0 to 1.The review describes AUC as scale-invariant and classification-threshold-invariant.
- Reviewed networks: Table III provides details of the deep networks used across the reviewed papers.