Neuroimaging Survey on Autism Spectrum Disorder Detection PDF

Exploring Learning Algorithm in Neuroimaging: A Comprehensive Survey on Autism Spectrum Disorder Detection Uday Singh*, Shailendra Shukla, Manoj Madhava Gore Department of Computer Science and Engineering, Motilal Nehru...

Exploring Learning Algorithm in Neuroimaging: A Comprehensive Survey on Autism Spectrum Disorder Detection Uday Singh*, Shailendra Shukla, Manoj Madhava Gore Department of Computer Science and Engineering, Motilal Nehru National Institute of Technology Allahabad, Prayagraj, 211004, Uttar Pradesh, India Abstract Neuroimaging captures the neuro activity and electrical disruptions in the human brain, which is crucial for diagnosing various brain disorders. Computer Aided Diagnosis (CADx) systems use neuroimaging to detect mental disorders like Alzheimer’s, Bipolar Disorder, Depression, Dementia, and Autism Spectrum Disorder (ASD) at an early stage. With the advent of Artificial Intelligence (AI), neuroimaging research has introduced innovative ways to analyze and interpret large amounts of data to classify different mental dis- orders. ASD is a neurodevelopmental condition characterized by repetitive and restricted behaviors, which create difficulties in engaging with other children. This survey presents a summary of recent studies involv- ing the application of neuroimaging methods, namely Electroencephalography (EEG), Structural Magnetic Resonance Imaging (s-fMRI), and Resting State Functional Magnetic Resonance Imaging (rs-fMRI), in the detection of ASD. The survey reviews current approaches and methods used for analyzing and interpreting neuroimaging data, as well as the challenges associated with these methods. Additionally, It focuses on various pre-processing techniques and neuroimaging ASD datasets. Overall, this survey aims to provide a comprehensive overview of recent advancements in utilizing AI-based neuroimaging techniques for ASD detection. Keywords: Neuroimage, Autism spectrum disorder, s-MRI, rs-fMRI, EEG. 1 Introduction Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by identified impairments in social interaction, communication abilities, and behavioral patterns. The symptoms and severity of ASD vary from child to child. It affects millions of children around the world.In 2020, the Centres for Disease Control and Prevention (CDC) said that 1 in 36 children in the US had ASD [3, 4]. Figure 1 illustrates the report on ASD prevalence from 2000 to 2020. This report states that ASD in children exhibited an upward trend. In the year 2000, 1 in 150 children were diagnosed with ASD. As of 2020, the prevalence has risen to 1 in 36 children with ASD. These statistics emphasize the pressing need for increased research and intervention to address the growing concern of ASD [3, 4]. Figure 1: 1 in X means one child from X is affected by ASD from 2000 to 2020. (Data source: Autism CDC ) 1  The Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) lists problems with social contact, speech, and limited and repetitive behaviors as signs of ASD. The DSM-5 merges the social and language deficits into a single characteristic. Now , DSM-5 cover only two characteristic for ASD.,,. Figure 2 shows the growing research on ASD detection from 2003 to 2021. It indicates that ASD research has grown in the last decade. Figure 2: The volume of research publications on autism detection has grown steadily from 2003 to 2021. This data was collected using an advanced search on Google Scholar, specifically targeting titles containing ”Autism or ASD” within the IEEE digital library, ACM library, Springer, PubMed, and Scopus. What are the challenges in accurately diagnosis and how can these challenges overcome? Diagnosing ASD is challenging due to the unavailability of laboratory-based diagnostic tests. Fur- thermore, there are no disease-modifying medications that target the core behavioral symptoms of ASD. Psychologists and specialists use various clinical tools to diagnose ASD. These clinical tools like , , , , , , and. These interview-based methods can be used for people of all age groups. The accuracy of ASD diagnosis relies on the physician’s proficiency, abilities, and availability. Ad- ditionally, interview-based methods have a drawback where the child’s family members might not provide truthful questionnaire responses, potentially leading to inaccurate diagnoses ,. Currently, there are no biological markers for ASD; therefore, behavioral observations and tests are used to diagnose the disorder. It may be difficult to obtain a diagnosis when access to specialists is limited, particularly in underdeveloped or rural regions. Late diagnoses can be problematic because it’s crucial to intervene early for the best results. There’s also a concern about overdiagnosis, where some children might be diagnosed with ASD when they have a different condition. Bias and misunderstanding about the disorder can also cause inaccurate or delayed diagnoses. Moreover, the limited understanding of ASD among healthcare professionals and the general public can also make it harder to diagnose accurately. Due to these limitations, researchers shifted to non-invasive techniques like medical imaging-based ASD detection. In recent decades, numerous studies have used medical imaging data for ASD detection ,. In the introductory section of this study, we answer three important questions that arise from the study’s overall scope. The initial question is: Why is medical imaging important for ASD detection? Medical imaging is a non-invasive method used for analyzing and intervening in medical cases by pro- ducing images of the internal aspects of the human body. A range of techniques is employed, including radiography , MRI , nuclear medicine , ultrasound , elastography , photoacoustic imaging , tomography , echocardiography , functional near-infrared spectroscopy , and magnetic parti- cle imaging. These techniques have become popular for detecting neuro disorder like ASD as it provides internal structural information that is not visible through external observation. A brain imaging technique is an experimental method that allows studying brain structure or function. These techniques meet specific criteria, including timing accuracy, spatial localization of cerebral function or structure, minimal invasiveness, repeatability, and suitability for ASD analysis. Structural MRI is widely used in neuroimaging for structural analysis, while functional imaging techniques such as EEG, PET, and fMRI are commonly used for functional analysis. However, EEG has been a reliable mapping method for the longest time. This survey covers Electroencephalography (EEG), s-MRI, and rs-fMRI modalities for ASD detection. What modality can be used in combination with learning algorithms to help ASD detetcion 2  The first modality, Electroencephalography (EEG) , records the brain’s electrical signals using scalp electrodes. These signals mirror the activity of brain cells, particularly pyramidal neurons. EEG lies in diagnosing and tracking neurological disorders by interpreting recorded electrical patterns. Researchers focus on EEG signals to learn about complex brain conditions for diagnoses and treatments to understand disorders such as Autism, Epilepsy, and Alzheimer’s. EEG analysis falls into four categories: time domain (sequencing events), frequency domain (identifying frequencies in signals), time-frequency domain (blending time and frequency analysis), and nonlinear methods (exploring complex relationships). Figure 3 illustrates EEG signals from a healthy child and an autistic child, showing consistent amplitude differences. Lower amplitudes in ASD signals suggest its presence. Figure 3: Comparison of EEG signals and their power spectra between ASD child and healthy. Specifically, it presents (a) EEG data from ASD, (b) the power spectrum derived from the EEG data of individuals with ASD, (c) an EEG data sample from typically developing individuals, and (d) the power spectrum obtained from the EEG sample of typically developing individuals The second modality is structural magnetic resonance imaging (s-MRI) , for ASD detection. In this technique, images of the brain’s structure are taken. It shows detailed anatomy and the difference between gray and white matter. Studies using s-MRI consistently found differences in the size of brain areas in children with ASD. Nevertheless, factors such as limited dataset size, diverse subject characteristics, and varying methodologies constrain the reliability and validity of these findings. Figure 4 shows the s-MRI of normal and ASD children. ASD child’s brain volume is 7 to 10% enlargement to the normal brain of a child ,. Figure 4: Normal vs ASD 3  The last modality for detecting ASD is rs-fMRI. It gives a detailed description of the brain’s activity. A study by Kelly and his team 2013 showed that rs-fMRI helped them understand how the brain’s cortex changes in children with ASD as they grow up. Improving the study and diagnosis of ASD involves focusing on specific affected brain areas. Detecting neurological markers at an early stage leads to accu- rate diagnoses and better care. Researchers have observed early brain changes when autistic behaviors and symptoms first emerge in infants. One important sign of brain development in autistic children is the rapid expansion of certain areas of the brain’s surface soon after birth [45, 46]. Moreover, it has been shown that patients diagnosed with autism exhibit larger ventricles due to reduced grey matter volume in certain locations, including the superior temporal cortex, thalamus, and striatum. Previous studies have also noted a reduction in cortical thickness during normal childhood brain development. AI is creating various Machine Learning (ML) and Deep Learning (DL) tools for solving complex problems such as prediction, simulating systems, and organizing information by simplifying them into categories. In preceding years, ML and DL models have been harnessed to identify Autism. Leveraging AI techniques and image processing, the Computer-Aided Diagnosis (CADx) system aids medical professionals in precisely detecting ASD. Figure 5 shows the basic model for neuroimage-based ASD detection. Figure 5: Flow of survey paper The presented survey offers a comprehensive and holistic exploration of ASD detection techniques, fo- cusing on diverse neuroimaging modalities and ML/DL methodologies. It emphasizes how important neu- roimaging data is to help understand the complex nature of ASD. The survey systematically categorizes and dissects ML approaches, ranging from classical methods to state-of-the-art DL models. The paper also evaluates different datasets used in ASD research, helping researchers find useful resources. It connects EEG, s-MRI, and fMRI methods, exploring their strengths and weaknesses for ASD diagnosis. Figure 7 illustrates the flow of our survey paper. This comprehensive survey comprises seven sections, offering an in-depth exploration of ASD research through neuroimaging techniques. Section 2 provides an overview of accessible preprocessing techniques and datasets dedicated to ASD studies, encompassing EEG, s-MRI, and f-MRI categories, along with their respective URLs. It also delves into the preprocessing methods applied to each category. Section 3 cov- ers methodologies related to feature extraction, selection, reduction, and machine/deep learning techniques within ASD detection, offering a broad perspective on neuroimaging feature extraction strategies, as well as methods for optimizing feature selection and reduction. Section 4 summarizes state-of-the-art procedures employed in ASD detection across EEG, s-MRI, and fMRI analyses. Insights from each modality are pre- sented within their respective subsections. Section 5 assesses future research directions and identifies research gaps in neuroimaging and ASD studies. Finally, Section 6 delivers concluding remarks, encapsulating the core findings and contributions discussed throughout the survey. 2 PRISMA Method This study was formed according to the systematic review guidelines outlined in Fig. 6 ,. This study examined around 1007 research papers from various well-known journals. Most publications were sourced from reputable academic publishers such as ACM, Springer, Elsevier, IEEE Library, Oxford Academic, and Frontiers. The papers are selected based on specific keywords (Autism prediction, Machine learning, autism Deep learning ASD, EEG-based autism detection, MRI-based autism diagnosis, Computer vision autism, Im- age processing ASD, Neuroimaging autism, Early autism detection, Behavioral markers, autism Biomarkers for ASD, Neurodevelopmental disorders detection, AI-based autism diagnosis, Data-driven autism detection, Autism spectrum disorder analysis, and Pattern recognition autism) related to using artificial intelligence for diagnosing ASD. These keywords encompassed various methods such as ML, DL, EEG data analysis, 4  Figure 6: flowchart for the selection and incorporation of relevant studies. MRI data analysis, computer vision, and image processing to diagnose and predict autism. We excluded papers that lacked a clear and transparent methodology, tool, strategy, or approach for the early diagnosis and prioritization of autism cases. In the initial phase of the Flowchart (Fig. 6), the search yielded 1007 papers that initially met the search criteria. After removing duplicates, 870 distinct articles remained for consideration during the screening phase. There are 690 article excluded in screening step. Subsequently, in the eligibility phase, 180 articles were deemed suitable for inclusion after a comprehensive examination of their full texts, while the remaining papers were excluded. Finally, in the last phase, a total of 82 papers were integrated into the study. Figure 7 illustrates the distribution of these papers over different years. Notably, the analysis focused more on papers published between 2017 and 2022, with a smaller number than before 2017. Figure 8 shows the basic model for neuroimage-based ASD detection. This model has three basic step 1) ASD neuroimaging dataset collection 2) Preprocessing technique and feature analysis. 3) ML/ DL Based ASD Classification. 3 Step 1- Neuroimaging Datasets of ASD This section provides an overview of the diverse and valuable neuroimaging datasets that are available for ASD research. These datasets play a crucial role in advancing our understanding of the neural underpinnings of ASD and contribute to the development of diagnostic and therapeutic strategies. The ASD dataset contains brain scan images which are used to identify autism. The main types of scans used in autism research are EEG, s-MRI), and rs-fMRI. Various datasets have been used to study and detect ASD, some of these datasets listed in Table 1. 5 Figure 7: Breakdown of papers reviewed for ASD detection using AI by year represented a percentage. 3.1 ABIDE-I ABIDE has collected data from several sites, resulting in bigger datasets than single-site research. This dataset encompasses both rs-fMRI and physical information obtained from 20 sites, with a composition of 1,112 sets of data. Out of these, 539 sets are from people with autism, and 573 sets are from individuals without autism. However, the dataset’s multi-site and multi-protocol nature introduces statistical noise and variations in rs-fMRI data. ABIDE provides phenotypic particulars, encompassing gender, age, scores from full IQ (FIQ) assessments, handedness, etc.. 3.2 ABIDE-II Besides ABIDE-I, ABIDE II comprises a network of 19 active sites, encompassing ten original charter institutions and seven newly incorporated members. These collaborating sites have contributed 1114 samples, covering 521 children diagnosed with ASD and 593 control participants, spanning an age range of 5 to 64 years. The dataset has been accessible to the scientific community since June 2016. Notably, all provided datasets have been meticulously anonymized to adhere to the stringent privacy standards set forth by the 1000 Functional Connectomes Project / INDI procedures and HIPAA regulations. 3.3 Autism BrainNet Autism BrainNet , functions under the Simons Foundation Autism Research Initiative (SFARI). This ini- tiative focuses on gathering and disseminating postmortem brain tissue obtained from individuals diagnosed with ASD and associated neurodevelopmental conditions. The purpose is to make this valuable resource accessible to eligible researchers globally. 3.4 Biobank Biobank are large repositories of biological samples and associated data used for medical research. These samples can include blood, urine, saliva, tissue, and other bodily fluids or materials and are collected from individuals who have informed consent to participate in research. The primary goal of a biobank is to support research into the causes, prevention, and treatment of various diseases and conditions. Biobank provide researchers with access to large and diverse collections of samples, identify biomarkers for disease diagnosis and prognosis, and develop new treatments and therapies. Biobank collect the Post-mortem brain and related bio specimens. Within the collection, there are 24 sets of rs-fMRI data, each accompanied by its corresponding s-MRI scans and phenotypic information. 3.5 NeuroVault NeuroVault is a publicly accessible repository for sharing neuroimaging data, particularly MRI datasets, which allows researchers to upload, share, and analyze neuroimaging data in a standardized format. The University of Texas at Austin maintains the database, and is freely accessible to the research community. This dataset comprises five distinct studies with participant counts of 277, 60, 50, 13, and 218 for each study, respectively. Within this database, a diverse array of data types are encompassed, including s-MRI, f-MRI, Tomography scans, and various others. 6  Figure 8: Block diagram of neuroimage based ASD detection 3.6 NDAR The National Database for Autism Research (NDAR) is a central repository for neuroimaging data related to ASD. The database contains a different MRI datasets, including structural, functional, and diffusion-weighted imaging data. It also provides tools for data analysis and visualization. The website for NDAR is , where researchers can access the data and learn more about the study protocols and 7  Table 1: Datasets Details URL Resource Type of Modalities Data Type Total Subjects ndar.nih. NDAR genetic, omics, NA 80,203 participants gov phenotypic, neu- roimaging sfari.org SFARI Characteristics re- Participant characteris- 3,000 participants from lated to observable tics, biological samples, the SSC dataset, 200 par- traits, brain imag- genomic information, ticipants from the Si- ing, and genetic at- brain imaging data mons VIP dataset, and tributes 50,000 participants from the SPARK dataset autism/ BioBank Brain tissue sam- NA More than 25 contribu- brainnet. for Autism ples obtained after tions have been made org BrainNet death along with starting from 2014. associated biologi- cal specimens. ABIDE- ABIDE Neuroimaging rs-fMRI, along with 539 participants were in- I,II s-MRI, and phenotypic volved in ABIDE I, while data ABIDE II included 487 participants. EEG AED Brain Imaging EEG data A total of 50 participants Database from Aus- tralia Brainmap. BrainMap maps representing Functional MRI (fMRI), There are 70 articles that org statistical informa- Positron Emission To- pertain to functional data tion mography (PET), and in the context of ASD structural coordinate- based findings neurovault. NeuroVault Statistical maps of Unprocessed statistical Five research investiga- org the human brain maps, parcellations, and tions included participant atlases resulting from groups of sizes 277, 60, MRI and PET studies 50, 13, and 218 in respec- tive studies umcd. USC Multi- Functional and NA 42 participants under- project. modal Con- structural con- went fMRI scans, while org nectivity nectivity matrices 51 participants under- Database derived from fMRI went DTI scans. provides and DTI data matrices representing brain con- nectivity datadryad. Dryad General repository long non-coding RNA Four separate studies org (lncRNA), magnetic consisted of participant resonance imaging counts of two, 34, 12, and (MRI), metabolic mark- 13 individuals, respec- ers (metabolites), and tively magnetoencephalography (MEG) neuro/ National BioBank Deceased brain tissue and There are 64 confirmed bio/bank. Institutes associated biological sam- cases of ASD and an ad- nih.gov of Health ples ditional 22 cases where Neuro ASD is suspected. BioBank (NBB) participant information. 8 3.7 EEG dataset The dataset containing EEG recordings was provided by the King Abdulaziz University located in Saudi Arabia. It contains 18 files: 8 from individuals with disorders and 10 from those without. The disorder group comprises 8 boys (ages 10-16), while the normal group includes 10 boys (ages 9-16). The EEG recordings were artifact-free, taken during rest at a 256 Hz sampling rate, using active electrodes and the BCI2000 system. Some other EEG dataset: The EEG-BNCI Horizon 2020 dataset encompasses EEG recordings sourced from both ASD and typical developed child. The EEG-TUH dataset contains electroencephalograph recordings obtained from autistic child. Access to the Temple University Autism BrainNet is facilitated via the dedicated website of the institution. 4 Step 2- Preprocessing technique and feature analysis 4.1 Preprocessing techniques for neuroimaging Neuroimaging analysis is difficult, time-consuming, and sophisticated. Additionally, artifacts and noise in the neuroimage make it difficult for radiologists to diagnose the disorders. Therefore, preprocessing is essential to refine the images before applying them to diagnosis. This section provides several pre-processing methods for neuroimaging data. Figure 8 show the various preprocessing technique of EEG, s-MRI, and fMRI. Figure 9: Preprocessing technique and software/tools 4.1.1 s-MRI pre-processing The preprocessing of s-MRI data is an initial step to ensure the quality and accuracy of subsequent analyses. This section outlines the procedures employed to correct artifacts, enhance tissue contrast, and standardize spatial dimensions. It deals with tasks like reducing noise, normalizing intensity, and aligning images for extracting valuable information about brain anatomy. Super resolution- Super-resolution , to enhance the resolution or quality of an image beyond its original limits. This is useful when you have low-resolution images and you want to generate higher- resolution versions. Super-resolution can be achieved through various methods, including single-image and multi-image. De-oblique- De-oblique preprocessing is a crucial step in s-MRI analysis. It refers to the process of removing any obliqueness or skewed orientation in data or images grid. Image data is gathered through oblique scans, deviating from the horizontal line to capture the entire brain. The ”3drefit” function in AFNI is used for skewed orientation problems. It restores alignment, enhancing subsequent analyses’ accuracy and image quality. Registration- Registration is aligning two or more images of the same scene, correcting for differences in position and orientation. It involves aligning and mapping different brain images together. This step is important for brain structure studies, as it establishes a standardized framework for subsequent investigations such as morphometry, localization, and connectivity analysis. Segmentation- The process of segmentation, as described by , refers to a preliminary stage in which brain tissues are categorized into three distinct types: grey matter, white matter, and cere- brospinal fluid. It enhances subsequent analyses by isolating these regions. Popular tools include FreeSurfer, which uses cortical and subcortical models to delineate structures, and FSL’s FAST, which 9  utilizes probabilistic estimation for tissue segmentation. Proper segmentation facilitates accurate struc- tural analysis and aids in understanding brain morphology and pathology. Inhomogeneity correction- Inhomogeneity correction method is used to compensate for signal intensity variations caused by scanner artifacts and tissue properties. It involves applying algorithms, like N3 or N4, to normalize image intensities across the brain. This step enhances data quality by ensuring consistent and reliable intensity information for improved brain tissue segmentation and spatial registration. Intensity standardization- Intensity standardization is used to normalize the intensity values of MRI scans across subjects. It adjusts for variations caused by different scanners or acquisition parameters, enhancing comparability. This involves scaling the intensity distribution of each scan to match a common reference distribution. It ensures that analyses are not confounded by intensity differences, promoting accurate and reliable results in inter-subject comparisons and group studies. Denoising- Denoising is used in image analysis. It reduce the noise and enhance image quality. It involves applying filters to remove unwanted variations while preserving essential features. In s-MRI, denoising techniques such as Gaussian filtering, wavelet threshold, or non-local means filtering are employed to improve image quality, enhance subsequent analyses, and minimize the impact of artifacts and scanner-related noise. Skull-stripping- Skull-stripping involves removing non-brain tissues such as the skull and scalp from images. It enhances subsequent analysis accuracy by isolating brain structures. FSL, Brain Ex- traction Tool (BET), and FreeSurfer recon-all use intensity gradients and geometric models to separate the brain from extraneous tissues, facilitating more accurate brain structure measurement and analysis. 4.1.2 f-MRI pre-processing technique The preprocessing of fMRI data is ensuring the quality and reliability of subsequent analyses. This section will outline the essential preprocessing steps aimed at mitigating noise artifacts and enhancing signal-to-noise ratios. The raw fMRI data will be subjected to motion correction, slice-timing correction, motion correction, and spatial normalization to a common template. Additionally, nuisance signal removal techniques will be applied to address physiological noise sources. These preprocessing steps collectively lay the foundation for meaningful and accurate insights into brain activity and connectivity. Removal of first N volumes- The initial stage in fMRI pre-processing entails the exclusion of the original n-volume pictures. This step is essential due to the behavior of hydrogen spins within the brain when exposed to a magnetic field. These spins take about 5 seconds to attain a stable state in the magnetic field’s orientation. To improve data quality and reduce noise, consider discarding the first n images to allow the signal to stabilize. Realignment (motion correction)- During fMRI image acquisition, subjects’ head movements can lead to positional variations in the brain images. Realignment is employed to rectify this issue by aligning all fMRI images to a designated reference image. This procedure mitigates the impact of motion-induced artifacts on the data. Consequently, the images share consistent coordinates across all voxels in the entire time series, ensuring uniform orientation. Typically, the initial image is chosen as the reference. Slice-time correction- Slice-time correction is a preprocessing technique used in f-MRI to correct for temporal differences in image acquisition between slices. During MRI scanning, slices of the brain are acquired sequentially, which can lead to distortions in the temporal order of neural activity. Slice- time correction adjusts each slice’s data to match the timing of a reference slice, ensuring accurate temporal alignment and enhancing the accuracy of subsequent analyses. Co-registration- Co-registration is a vital neuroimaging technique that aligns multiple images from the same individual, correcting for minor motion or spatial variations. Used in studies like fMRI and PET, it ensures accurate comparison by registering images to a common space. Co-registration en- hances spatial accuracy, reduces artifacts, and enables precise analyses of brain structure and function, aiding in understanding neurological processes and disorders. Spatial smoothing- Spatial smoothing is a technique used in neuroimaging to enhance signal- to-noise ratio and reduce noise-related variability. By averaging neighboring voxel values, spatial smoothing reduces minor image inconsistencies caused by measurement errors. However, excessive smoothing can blur fine details. Gaussian kernels are commonly applied during preprocessing to strike a balance between noise reduction and preserving meaningful anatomical and functional information in the data. Temporal filtering-Temporal filtering is a technique used in the analysis of f-MRI data. This method focuses on enhancing meaningful signals and reducing unwanted noise in the time series data collected during scans. Apply filters (e.g., low-pass for high-frequency noise, high-pass for slow fluctu- ations) to enhance data quality. This enhancement in data quality results in a better signal-to-noise ratio, aiding in detecting neural activity patterns linked to cognitive functions. 10 4.2 Software/tools for preprocessing To prepare the neuroimaging scans before analysis, there are many software options accessible to both radiologists and researchers. Notable packages include SPM, AFNI, FreeSurfer, FSL, REST, and DPARSF. These tools play a significant role in examining brain images within neuroimaging studies. Here’s a concise introduction to each of these tools: SPM (Statistical Parametric Mapping)- SPM is a MATLAB-based software tool that is widely used for the analysis of functional and structural brain images. It provides a range of functions for tasks such as preprocessing, statistical analysis, and visualization. AFNI- AFNI is a software tool that is widely used for the analysis of functional brain images. It provides a range of functions for tasks such as preprocessing, statistical analysis, and visualization. AFNI is written in C and can be used on multiple platforms. FreeSurfer- FreeSurfer is a software tool for analyzing structural brain images. It provides a range of functions for tasks such as preprocessing, segmentation, surface reconstruction, and visualization. FreeSurfer is written in C and can be used on multiple platforms. FSL (FMRIB Software Library)- FSL is a software tool for the analysis of both functional and structural brain images. It provides a range of functions for tasks such as preprocessing, statistical analysis, and visualization. FSL is written in C++ and can be used on multiple platforms. REST- REST is a software tool for the analysis of brain images. It provides a range of functions for tasks such as preprocessing, statistical analysis, and visualization. REST is written in MATLAB and can be used on multiple platforms. DPARSF- DPARSF is a software tool for the analysis of resting-state functional brain images. It provides a range of functions for tasks such as preprocessing, statistical analysis, and visualization. DPARSF is based on MATLAB and can be used on multiple platforms. 4.3 Feature analysis for ASD diagnosis The subsequent steps involve feature extraction, selection, and reduction, aimed at distilling essential infor- mation from the neuroimaging data. Subsequently, a range of ML and DL techniques are applied to leverage this data for accurate ASD diagnosis. These methods utilize the power of advanced algorithms to decipher intricate patterns that might not be discernible through conventional means. Figure 9 shows the feature extraction, selection, and reduction methods. 4.3.1 Feature extraction techniques In the initial stages of model development, the process of feature extraction is fundamental. This step involves different techniques for deriving a fresh set of features beyond the ones already present. It’s essential to choose significant features from an extensive pool in order to enhance the model’s effectiveness. Several methods for feature extraction exist, each serving to extract pertinent information from a dataset. These techniques encompass graph-based, texture-based, non-linear, connectivity matrix-based, and statistical measures-based methods. The following section provides a brief overview of these methods. 11  Figure 10: Feature extraction, selection and reduction technique Voxel-based morphometry (VBM): VBM is a technique that involves comparing the volume and density of gray and white matter in different brain regions. VBM is often used in studies of neurodegenerative diseases and other conditions that affect brain structure. Functional connectivity: Functional connectivity feature extraction methods involve quanti- fying the interactions and relationships between different brain regions based on neuroimaging data. One common approach is seed-based correlation, where a seed region’s activity is correlated with other regions. Independent Component Analysis (ICA) identifies independent networks. Graph Theory measures, like node degree, assess node importance. Diffusion tensor imaging (DTI): DTI is harnessed for its ability to capture white matter microstructural characteristics. Various feature extraction techniques are employed to derive valuable information from DTI data, such as Fractional Anisotropy (FA), and Radial Diffusivity (RD). These measures provide insights into the integrity and orientation of neural pathways, facilitating a deeper understanding of brain connectivity alterations associated with conditions like ASD. Graph theory Graph theory is a mathematical framework used to study the organization of complex networks of brain. Graph theory is used to identify important hubs in the brain network and to understand how information is transmitted between different brain regions. Independent component analysis (ICA): ICA is a method that separates a complex signal into its underlying independent components. ICA is often used in functional MRI studies to identify patterns of brain activity associated with different cognitive processes. Region of interest (ROI) analysis: ROI analysis involves selecting specific brain regions and analyzing the activity or structure within those regions. ROI analysis is often used to test specific hypotheses about the function of different brain regions. (MVPA): Multivariate pattern analysis is a method that analyzes patterns of brain activity across multiple voxels or brain regions. MVPA is used to decode different cognitive states or to identify patterns of brain activity that are associated with specific tasks or stimuli. Structural equation modeling (SEM): SEM is a statistical method used to test causal hy- potheses about the relationships between different brain regions. SEM is often used to understand the functional organization of the brain and to identify the key drivers of cognitive processes. 4.3.2 Feature selection/reduction methods A collection of data includes various characteristics, and certain characteristics might not hold importance in identifying a disease. By using suitable techniques to choose the most pertinent characteristics and decrease their number, the accuracy of disease diagnosis can be enhanced. The following are essential approaches for selecting and reducing characteristics. 4.3.3 Feature selection methods Feature selection involves choosing the most relevant attributes to train a model. Feature selection method can be broadly categorized into two main types: (1) supervised, (2) unsupervised methods. 12 Supervised feature selection methods like Fisher, Relief, Correlation, and Chi-Squared select attributes based on how they differentiate between different classes. For example, Fisher chooses attributes with minimal inter-class distance and maximized intra-class distance. Relief assesses relevance by comparing attributes among neighboring and selected samples. Correlation methods measure the connection between features, often used in various diagnoses. Chi-squared identifies features dependent on input data. Feature selection in unsupervised techniques involves exploring methods like scatter ratio, Kernel method, variance, Ranking methods, agglomerative hierarchical clustering, and random projection.. These approaches offer diverse avenues for feature choice. In the realm of feature selection techniques, several are extensively adopted, including: Filter Methods Filter methods are a category of feature selection techniques that employ statistical tests to assess the relevance of each feature independently of the target variable. These methods rank features based on their scores and then select the top-ranked features for further model construction. Examples of filter methods include mutual information, chi-squared tests, and correlation, among others ,. Wrapper Methods:Wrapper methods are evaluating subsets of features using a chosen machine learning algorithm. These subsets are treated as different ”wrappers” around the algorithm. The performance of the algorithm is measured for each subset, allowing the selection of the subset that produces the best results. While more computationally intensive, wrapper methods can provide better feature combinations for improved model performance. Embedded Methods:Embedded methods are feature selection techniques integrated within the pro- cess of training a machine learning model. These methods automatically assess feature importance while the model is being built, enhancing its performance by focusing on the most relevant features. This integration helps the model learn which features contribute most effectively to its predictive accuracy, leading to improved efficiency and generalization ,. Dimensionality Reduction Methods Dimensionality reduction methods aim to simplify complex data by reducing the number of variables while preserving meaningful information. These methods minimize the dimensions of multidimensional data, improving visualization, analysis, and processing in ML/DL data analysis. The purpose of the model, the number of available features, and the extent of the dataset all influence the choice of feature selection technique. It is important to select the right feature selection method to avoid overfitting and improve the model’s performance. 4.3.4 Feature reduction methods Feature reduction, a pivotal preliminary step in constructing a model, involves transforming high- dimensional data into a more concise, lower-dimensional representation. One prominent method for accomplishing this is Principal Component Analysis (PCA). Various techniques exist for feature reduction, encompassing: – PCA: PCA, or Principal Component Analysis, is an incredibly powerful technique employed in data analysis and pattern recognition. It is a process that collects interrelated variables and converts them into a new set of independent variables known as principal components. These components capture the most significant information in the data, reducing complexity while re- taining essential patterns. PCA is widely employed in fields such as image processing, genetics, and finance for simplifying data while minimizing loss of information. – Recursive Feature Elimination RFE: RFE or Recursive Feature Elimination involves iter- atively training a model on the full feature set and then removing the least important feature. This process is repeated until a specified number of features remains. RFE helps to identify the most relevant features for the model, enhancing its efficiency and interpretability by discarding less influential attributes. – Linear Discriminant Analysis (LDA) LDA is a statistical technique that analyzes and classifies data by finding linear combinations of features that best separate multiple classes or groups. It proves particularly valuable for tasks such as pattern recognition and data categorization. LDA’s primary goal is to maximize the variance ratio between different classes to that within individual classes. This optimization ensures that data points are distinctly separable, resulting in more effective classification and pattern recognition.. – t-SNE: t-SNE is a dimensionality reduction technique used in machine learning and data visual- ization. Its goal is to reduce the dimensions used to store data without losing information about the local interactions between data items. t-SNE emphasizes the clustering of similar data points, making it particularly useful for revealing patterns and structures in complex datasets. – Autoencoders:These represent neural network models that acquire the ability to condense the input data into a space with fewer dimensions and subsequently restore it to its initial structure. Autoencoders are used for unsupervised feature reduction, and they work by learning a compressed representation of the input data while preserving its essential features. Overall, the choice of feature reduction method depends on the nature of the data, the research question, and the available computational resources. 13 5 Exiting Modality for ASD Detection 5.1 Detection of Autism using EEG signals In the domain of ASD detection, Electroencephalography (EEG) has emerged as a potent tool for unraveling the intricacies of neural activity and connectivity patterns. EEG, a non-invasive technique, holds the potential to provide valuable insights into the underlying neurophysiological alterations as- sociated with ASD. By capturing the real-time electrical activity of the brain, EEG data offers a window into the dynamic processes that contribute to the manifestation of ASD traits. This section of the survey delves into the landscape of EEG-based ASD detection methodologies, exploring the di- verse range of approaches harnessed to decode the enigmatic complexities of ASD using neural signals. Through a comprehensive review, we shed light on the progress, challenges, and innovative avenues that EEG-based research offers in enhancing our understanding of ASD and its early diagnosis. M. Hashemian and H. Pourghassem examine the challenges involved in diagnosing ASD and the potential of EEG to address these challenges. It reviews studies on EEG biomarkers, EEG patterns, and machine learning-based approaches for ASD diagnosis. It concludes that EEG analysis shows promise but needs more research to determine its clinical usefulness. The authors suggest future research should focus on developing more reliable EEG biomarkers for ASD, improving machine learning-based approaches, and exploring combining EEG analysis with other diagnostic tools. Fahd A. Alturki et al. explored how the Common Spatial Pattern (CSP) method can help diagnose autism and epilepsy using EEG signals. They collected EEG data from 80 participants and applied CSP to find important patterns in each group. Using machine learning, they categorized participants by their conditions using these patterns. The results showed that combining CSP with band LBP provided the most accurate outcomes. Specifically, the CSP-LBP-KNN combination achieved classification accuracy of 98.466% for autism and 98.62% for epilepsy diagnosis. Bogdan Alexandru Cociu et al. presented a comprehensive method to explore the brain mecha- nisms of ASD. They combined data from EEG, fMRI, and DTI scans to gain insights. Their approach involved pinpointing brain sources using EEG data, aligning them with fMRI-detected active brain areas, and then examining the connectivity between these regions. The strength of this connection was found to relate to the physical characteristics of brain structures obtained from DTI analysis. William Bosl et al. identified early biomarkers for neurodevelopmental disorders using modified multiscale entropy (mMSE) analysis of resting state EEG signals. It successfully distinguishes typically developing children from high-risk infants for autism. The mMSE trajectory differs in high-risk infants, particularly at 9-12 months, achieving over 80% accuracy in classifying groups at 9 months, especially for boys. This underscores mMSE’s potential as a cognitive development and autism risk biomarker in infants. William J. Bosl et al. collected EEG data from 99 infants with ASD-sibling risk and 89 typically developed children, ages 3 to 36 months. The nonlinear EEG features were used for accurate ASD prediction, achieving over 95% specificity, sensitivity, and PPV at times and correlating with ADOS severity scores, indicating potential early digital biomarkers. Noura Alotaibi, Koushik Maharatna aimed to develop early autism diagnosis strategies using EEG-based functional brain connectivity. Graph parameters are extracted from phase-based networks. They used a cubic SVM model that achieved 95.8% accuracy, demonstrating its potential for ASD classification and identifying connectivity differences in theta band. Jiannan Kang et al. enrolled 97 children (ages 3-6), utilizing EEG and eye-tracking data for autism identification via SVM. Com- bining modalities achieved 85.44% accuracy and AUC of 0.93 with 32 features. This approach proves promising for aiding ASD diagnosis in children, enhancing diagnostic processes through comprehensive data integration. M Radhakrishnan et al. evaluated deep convolutional architectures for ASD detection using EEG. ResNet50 showed the highest accuracy (81%) through stratified K-fold cross- validation, outperforming other models like AlexNet and VGGNet. ResNet’s skip connections aided accurate feature extraction. Thus, ResNet50 is a promising method for automatic ASD detection from input spectrograms. Thirumal, S.; Thangakumar, J. proposed a novel Greedy RFD feature selection and fusion of EEG- behavioral features. Optimized GBM with Greedy RFD enhances classification accuracy compared to other methods. Fused data improves ASD classification. Md. Nurul Ahad Tawhid et al. presents a diagnostic framework for ASD using EEG signals. Raw EEG data is pre-processed, transformed into spectrogram images, and evaluated by ML/DL models. The DL model achieves 99.15% accuracy, surpassing the ML model’s 95.25%. This approach offers the potential for automatic ASD diagnosis and biomarker discovery, aiding future computer-aided diagnosis systems. Avirath Sundaresan et al. investigate stress assessment using EEG signals, comparing SVM and deep learning methods. 11 subject-dependent models, including conventional BCI and deep learning, are trained on neurotypical and ASD participants’ EEG during stress induction tasks. A two-layer LSTM RNN achieves 93.27% accuracy, showing potential for real-time stress assessment and mitigation, benefiting both populations. Enzo Grossi et al. utilized limited EEG data from C3 and C4 channels with advanced machine learning (ML) techniques to distinguish ASD in children. Employing a hybrid system named TWIST, combining an evolutionary algorithm and neural network, yielded 100% accuracy in identifying ASD cases from controls in study 1 and 94.95% accuracy in differentiating neuropsychiatric disorders in 14 study 2. These findings imply the potential of using ML on even small EEG segments for early autism detection in newborns. Mehmet Baygin et al. proposed a CAD model for ASD detection using EEG signals. For signal-to-image conversion, a combination deep lightweight feature extractor uses D-LBP and STFT. Pre-trained models (MobileNetV2, ShuffleNet, SqueezeNet) extract deep features refined by ReliefF. SVM classifier achieves 96.44% accuracy, highlighting its suitability for aiding autism diagnosis in medical settings. Nur Fadzilah Harun and colleagues devised an innovative approach employing EEG signals and advanced ML such as ANN and SVM, to efficiently detect ASD in children. Their method significantly boosts the precision and efficiency of ASD diagnosis by positively identifying distinct brainwave patterns and emotional states. In a separate investigation, Jing Fan and her team developed a driving simulator using virtual reality technology for the purpose of training autistic children. They utilized EEG signals obtained from a 14-channel neuroheadset, combined with therapist evaluations, to build classification models such as Bayes networks and SVM. The outcomes showcased an accuracy of over 80% in categorizing engagement and mental workload, along with a classification accuracy of over 75% for emotional states. This underscores the potential of employing an adaptable VR training system for interventions related to autism. Enzo Grossi, Chiara Olivieri, and Massimo Buscema utilized the MS-ROM/I-FAST technique to differentiate individuals with ASD from typically developing ones, achieving 100% classification accu- racy using EEG data and machine learning, showcasing its potential for non-invasive ASD diagnosis. Alvin Sahroni et al. assessed brain abnormalities in autism using spontaneous EEG data from sleeping subjects aged 3-10. Fast Fourier transform extracted delta, theta, alpha, beta, and gamma band powers. The autism group exhibited elevated alpha activity p < 0.05, predominantly in frontal regions of both hemispheres p < 0.05. Similar abnormalities were found during awake and resting conditions. In the study by Bosl et al. , a combination of windowing, filtering processes, and a RNN-GRU was employed to analyze the multi-scale entropy of resting-state EEG. Their approach included the use of classifiers such as SVM, KNN, and naive Bayesian, yielding an impressive 80% accuracy for ASD classification. Stahl et al. investigated ERP of direct gaze trials in infants’ sib- lings using ICD-10 criteria. Their research involved the application of regularized discriminant function analysis and SVM, leading to a classification accuracy of 64% for ASD. Cohen et al. investigated neonatal EEG data, mainly focusing on abnormalities in auditory brainstem response (ABR). Their analysis incorporated piece-wise linear regression and ROC analysis, resulting in an accuracy of 85% for ASD identification. Miron et al. based their study on DSM-IV criteria, specifically assessing ABR latencies. They determined optimal latency values using ROC analysis with a leave-one-out procedure, achieving a sensitivity of 70% for ASD detection. Lastly, Gabard et al. undertook a longitudinal EEG study, evaluating power parameters using the ADOS assessment. Their application of logistic regression models yielded a positive predictive value (PPV) of 72%, sensitivity (SEN) of 81.82%, and specificity (SPE) of 82.35% for ASD classification. In combination, these studies collectively demon- strate diverse EEG analysis techniques and classification models for accurate ASD identification across various populations using a wide array of EEG features [105, 117–120]. Numerous EEG-based methods for ASD detection have emerged, capitalizing on the potential of EEG to reveal neural connectivity patterns. Researchers employ diverse strategies, such as biomarker identifi- cation, machine learning integration, and neurobiological exploration. Some methods exhibit promising diagnostic accuracy, distinguishing ASD and typically developing individuals. Notably, several stud- ies emphasize the significance of specific EEG frequency bands, particularly alpha activity in frontal brain regions, as indicative of ASD-related abnormalities. Despite advancements, challenges like ro- bust biomarker establishment and reliable classification persist. Integrating EEG with other modalities, such as fMRI and eye-tracking, adds depth to ASD understanding. While these approaches collectively enhance early diagnosis prospects, the field demands further research to establish EEG’s clinical utility and improve accuracy. However, EEG-based ASD detection encounters limitations. Methodological variations across studies and small sample sizes affect generalizability. Interpretation complexity arises from diverse EEG pat- terns, requiring refined algorithms for precise classification. Despite notable accuracy, challenges persist in achieving consistent results across different age groups and developmental stages. The reliability of EEG biomarkers remains a concern, necessitating additional validation. Integrating EEG with other modalities introduces complexity and demands robust data fusion techniques. Furthermore, real-world clinical applicability requires addressing practical challenges like participant cooperation and external noise. While EEG-based ASD detection holds promise, future research must focus on standardization, cross-validation, and extensive clinical trials to validate its potential as a reliable diagnostic tool. 15  Table 2: Summary of EEG Based articles for ASD classification. References Dataset Participants Preprocessing/Filters Feature Extrac- Classification Performance tion/Selection Method criteria (%) Ref. EEG 80 child CSP and band CSP-LBP-KNN KNN Accuracy= LBP 98.46% and 98.62% for ASD Ref. EEG 12 ASD child Used the trial aver- NA Cubic Support vec- Accuracy= and 12 normal aged phase locking tor machine 95.8% , children value and Speci- ficity= 92% Ref. EEG and 97 children band-pass filter be- Minimum SVM Accuracy= eye track- aged from 3 to tween 0.5 and 45 redundancy- 85.44% ing data 6 years Hz maximum rele- vance (MRMR) Ref. Natus 10 TD children 22 channels, 500 NA ResNet50 Accuracy= Nihon and 10 ASD Hz sampling fre- 81% Ohden children quency with low- EEG pass and high-pass filters at [0.53, 70] Hz.. Ref. EEG NA Wavelet Trans- Correlation , Infor- K-Nearest Neigh- NA form, dimension- mation Gain bour (KNN), Min- ality reduction imum Redundancy , irrelevant data Maximum Rele- removal vance (MRMR) and the. Ref. King Ab- ASD: 3 girls, Artifacts were re- Ten-fold cross- NB, LDA, RF, accuracy= dulaziz 9 boys, HC: 4 moved using re- validation kNN, LR and 99.15% University boys. referencing, filter- SVM (KAU) ing, and normaliza- Dataset tion Ref. EEG 5 HC and 8 High-pass filters NA LSTM-RNN and accuracy= Dataset ASD. erase slow trends SVM 93.27% and 50 Hz low-pass filters remove line noise. Ref. EEG data Study 1 in- The raw EEG NA Sine-net ANN, Study 1 cluded 15 teens time-series were TWIST, and accuracy= with ASD and processed using a BackPropagation 100%,Study Study 2 had 20 feature extraction accuracy= with additional algorithm, result- 94.95% neuropsychi- ing in 794 quan- atric titative features extracted using the TSFRESH. Ref. EEG sig- 16 child(3–10 delta, theta, alpha, NA NA autism nals years old) beta, and gamma), group Fast Fourier trans- exhibited form elevated alpha activity p < 0.05 Continued on next page 16  Table 2 – Continued from previous page References Dataset Participants Preprocessing/Filters Feature Extrac- Classification Performance tion/Selection Method criteria (%) Ref. EEG 122 child 1D-LBP and NA Decision Tree, SVM Ac- STFT Discriminant curacy Analysis, LR, =96.44% SVM, kNN Ref. EEG sig- child age 4 to Emotiv test bench NA KNN and SVM ANN Ac- nals 12 software curacy= 90.46% and SVM accuracy 88.62% Ref. EEG sig- 16 male age 13 0.2–45Hz band- hanning window naı̈ve Bayes, SVM, Accuracy nals to 18 years pass filter function MLP, KNN, ran- for mental dom forest, and workload J48. = 80% and emo- tional accuracy =75% for ASD Ref. EEG 8 children with Filtering, ICA, TensorFlow DCNN Accuracy ASD and 12 Segmentation, and = 80% for Epilepsy 18 PSDED ASD with HC Ref. King Ab- 9 ASD 10 HC Windowing, Filter- Keras, TensorFlow 2D-CNN + Soft- Accuracy dulaziz ing Process, RNN- max = 99.5% University GRU, FastICA for ASD (KAU) Dataset Ref. EEG HR = 46; con- Windowing, Filter- Multi-scale en- kNN, SVM and Accuracy trol = 33 ing Process, RNN- tropy of rs-EEG naıve Bayesian = 80% for GRU, FastICA ASD Ref. EEG HR = 19; con- ICD-10 direct gaze trials SVM Accuracy trol= 17 SVM = 64% for ASD Ref. EEG Normal IABR and AMA NA Piece-wise linear Accuracy ABR=28 slope regression, ROC = 85% for and abnormal ASD ABR=46 Ref. EEG ASD=30; con- DSM ABR latencies Optimal latency Sensitivity trol=30 value based = 70% for on ROC with ASD leaveone-out pro- cedure Ref. EEG Short HR ra- ADOS Longitudinal EEG ROC of logistic re- PPV=72%, tios: 31 (ASD), power parameters gression models SEN=81.82%, 71 (N), and 69 SPE=82.35%; (LR). 5.2 Diagnosis of Autism using s-MRI modality S-MRI can diagnose ASD. This imaging method shows brain structure and can distinguish ASD from typically developing people. ASD patients have different prefrontal cortex, temporal lobe, cerebellar volume, thickness, and surface area. S-MRI can also identify ASD risk in six-month-olds, enabling early intervention and therapy. s-MRI can potentially diagnose ASD. However, further study is needed 17 to create accurate biomarkers. Akshoomoff et al. explored whether early childhood MRI brain measurements can differentiate ASD from typical development and correlate with functioning. They quantified brain volumes in 52 ASD boys and 15 typical children. The MRI data accurately classified 95.8% of ASD and 92.3% of controls, highlighting its diagnostic potential. Additionally, cerebellar and cerebral size variations were tied to diagnostic and functional outcomes in young ASD children. Furthermore, Marek D Shen et al. aimed to identify early brain anomalies in infants who later develop ASD. Magnetic resonance imaging was utilized on 6–9 month-old infants, with longitudinal scans conducted at 6–9, 12–15, and 18–24 months.Infants later diagnosed with ASD had larger brains and more brain fluid, regardless of their initial risk level. These findings underscore the potential of MRI markers for early ASD detection, offering a novel approach to identifying at-risk children. Shen et al. conducted studies to enhance ASD detection through neuroimaging and machine learning approaches. In their initial work , they found that more extra brain fluid (EA-CSF) at six months is connected to ASD development, persisting until 24 months. More severe autism symptoms were linked to greater EA-CSF volume. Using this information, they developed a predictive model that could diagnose ASD with 69% accuracy. In a subsequent study , they confirmed elevated CSF volume in high-risk infants with ASD aged 2-4 years, preceding behavioral symptoms. Their machine learning-based approach predicted ASD diagnosis with higher accuracy (83%), and Sleep problems and decreased non-verbal skills were linked to higher CSF volume. In another investigation , the team focused on machine learning utilizing anatomical brain features for ASD diagnosis. Among children aged 18-37 months, their model based on regional average cortical thickness outperformed traditional volume and surface area-based methods. This technique holds promise for providing stable biomarkers for supplementary ASD diagnosis, demonstrating the significance of machine learning, especially in datasets with limited samples. Overall, Shen et al.’s work highlights the potential of innovative imaging and computational methods in advancing early ASD detection and understanding its underlying mechanisms. Gaunnan et al. suggested a new method to detect ASD early. They used a deep-learning technique called dilated-dense U-Net to analyze brain scans of infants around 24 months old and identify specific brain regions like the amygdala and hippocampal subfields. By addressing low tissue contrast, their method identifies overgrowths in the amygdala and hippocampus subfields (CA1-3) using the NDAR dataset. This insight hints at potential links to ASD emergence, demonstrating improved segmentation accuracy and providing valuable insights into ASD development. In a related study, Hazlett et al. delve into ASD detection using neuroimaging. The findings demonstrate early neurodevelopmental alterations in ASD, indicating a notable augmentation in cortical surface area expansion during the 6-12 months before the later surge in brain volume growth observed between 12-24 months in infants with a heightened susceptibility to autism. This volume overgrowth correlated with emerging autistic behaviors and was accurately predicted at 24 months using a deep-learning algorithm with 81% positive predictive value and 88% sensitivity. Ismail et al. presented a novel computer-aided diagnosis (CAD) system for autism detection using MR brain images. The proposed approach integrates shape data extracted from the cerebral cortex (Cx) and cerebral white matter (CWM) to improve the resilience of the classification process. A 3D joint model segments Cx and CWM, followed by Spherical Harmonic analysis to derive metrics. The CAD system employs a multi-level deep network to fuse features and diagnose. Evaluated on ABIDE (8–12.8 years) and NDAR/Pitt (16–51 years) databases, achieving 93% and 97% accuracy respectively. Tested on NDAR/IBIS database for infants, yielding 85% accuracy, indicating potential for early diagnosis. Kong et al. proposed ASD detection using individual brain networks and a DNN classifier on the ABIDE I dataset. Networks capture ROI connectivity. The top 3000 features, chosen by F-score, are input to the DNN, achieving 90.39% accuracy and 0.9738 AUC in ten-fold cross-validation. It surpasses some methods, enhancing ASD classification via connectivity analysis. Pinaya et al. introduced an innovative approach for investigating brain-based disorders, addressing concerns about data requirements and interpretability. It employs a ”deep autoencoder” neural network to establish a normative model using structural MRI data from 1,113 healthy subjects. This model is then applied to assess neuroanatomical deviations in 263 patients with schizophrenia and ASD, generating distinct deviation values for each condition (p ¡.005). The method reveals disease- specific neuroanatomical patterns, consistent with existing literature. The deep autoencoder proves versatile for detecting deviations in neuropsychiatric populations, suggesting broader applications in different neuroimaging modalities and preprocessing methods. Sujit et al. employed an ensemble deep learning (DL) method to evaluate brain MRI image quality for autism detection. It employs cascaded deep convolutional neural networks (DCNNs) on axial, coronal, and sagittal MRI planes. Using 1064 autism-control ABIDE database images and 110 multiple sclerosis images, the research reports AUC, sensitivity, specificity, accuracy, PPV, and NPV. The ensemble DL model demonstrates AUC 0.90 (ABIDE) and 0.71 (CombiRx), signifying accurate image quality assessment for multicenter structural brain MRI studies. Shahamat et al. employed a 3D-CNN for classifying MRI scans. A GABM technique is introduced to visualize the 3D-CNN’s functional aspects. The method involves training the CNN on preprocessed MRI scans and applying a genetic algorithm to identify brain regions crucial for classification. Evaluation using ADNI (140 samples) and ABIDE (1000 samples) datasets achieved accuracies of 0.85 and 0.70, respectively. GABM revealed 6-65 knowledgeable brain regions 18  (ADNI) and 15-75 (ABIDE). The method improved classifier performance and interpretability. Future work could explore multiple brain atlases for better region identification and extend the approach to other neuroimaging data. Gingjing Gao et al. presented a novel approach for ASD detection using structural MRI data. The study combines a convolutional neural network (CNN) with individual structural covariance networks. They utilize the ABIDE consortium dataset. The proposed method outperforms existing approaches, achieving 71.8% accuracy across sites. Discriminative features are identified in the prefrontal cortex and cerebellum, potentially serving as early ASD biomarkers. The CNN-based framework proves effective for ASD diagnosis via individual structural covariance networks. Mingli Zhang et al. proposed an ASD detection method using fused T1 MRI and fMRI images. Combining discriminative learning and CNNs creates imaging-based ASD biomarkers. Tested on an 1127-subject, 34-site database, the approach shows competitive results, capturing ASD patterns. Table 3: Summary of s-MRI Based articles for ASD classification. References Dataset Participants Preprocessing Feature Extrac- Classification Performance tion/Selection Method criteria (%) Ref. s-MRI TD-15; provi- ADOS Volumes of gray Discriminant func- Accuracy(ASD) sional autism- and white matter tional analysis =95.8%; 52; ASD-42, in the cerebellum, Accuracy PDDNOS-10 as well as the sizes (Control) of the anterior and =92%3 posterior vermis regions. Ref. s-MRI HR=33; DSM-IVTR and Extra-axial CSF, Discriminant func- Accuracy= LR=22 (ASD: ADOS brain volume tional analysis 76% Sen- 10) sitiv- ity=78%, Speci- ficity=79% Ref. s- MRI HR=221 (HR- ADOS Extra-axial CSF RUSBoost trees Accuracy=69%, ASD=47,HR- volume and cere- Sensitiv- N=174);LR=122 bral volume ity=0.66, (all non-ASD) Speci- ficity=0.68 Ref. s- MRI ASD=159; ADI-R and ADOS NA Discriminant func- PPV=83%; TD=77 method tional analysis Accu- racy=78%; Sensitiv- ity=84%; Speci- ficity=65% Ref. s- MRI ASD(46) and DSM -IV Cortical thickness Random forest, Accuracy TD(39) chil- naıve Bayes and =80.9%, dren SVM Sensitiv- ity=81.3%, Speci- ficity=81%, AUC=88%, MCC=62% Ref. s- MRI HR=106 (HR- ADI-R Cortical surface DL method PPV=81%, ASD=15; HR- area Sensitiv- N=91); LR=42 ity=88%, Speci- ficity=95% Ref. NADAR 60 child with In house tools 3D Patches Ex- DDUNET NA ASD, 211 child traction with HC Continued on next page 19  Table 3 – Continued from previous page References Dataset Participants Preprocessing Feature Extrac- Classification Performance tion/Selection Method criteria (%) Ref. NADAR total (47 ASD FSL and iBEAT Segmentation, Softmax Accuracy pitt and and 47 HC) Shape Feature = 97% ABIDE- Extraction 1,NDAR IBIS Ref. ABIDE-I 104 HC and 78 FreeSurfe and De- Construction of In- Sparse-AE Accuracy=90.39% ASD strieux dividual Network Sen=84.37% and F-score Speci- ficity=95.88% Ref. HCP (1113 HC and FreeSurfe Normalization DEA and Softmax Accuracy=63.9% ABIDE-I 78 ASD) 104 method and HC One-Hot Coding method Ref. CombiRx 1112 SPM12 32 Slices Along DCNN,Softmax Accuracy= and Each Axial, Coro- 84% ABIDE nal, and Sagitta Ref. (ABIDE- NA FSL Chromosome 3D-CNN + Soft- NA II) Encoding max Ref. ABIDE-I NA FSL NA 3D-CNN, Softmax NA Ref. ABIDE-I 567HC and SRI24 Morpho-logical 2D-CNN (ResNet) Accuracy 518ASD Co-variance = 71.8% method Ref. IMPAC 537child with sMRI Measures of NA CNN + Softmax Accuracy ASD and 590 Cortical Thickness, = 69% with HC Surface Area 5.3 Diagnosis of Autism using fMRI modality f-MRI is employed to investigate the patterns of brain activity and connections among individuals with ASD. fMRI detects neural activity by measuring brain blood oxygen levels. ASD’s verbal, sensory, and social cognition have been studied using fMRI. fMRI studies show that ASD patients have different brain activity and connectivity than normally developing people. These distinctions may help diagnose and treat ASD by revealing its brain processes. Xiaoxiao Li et al. proposed a two-step approach to identify ASD utilizing fMRI data. Their method employs advanced DNN to differentiate between ASD child and those without, while also analyzing the significant features highlighted by the classifier. This approach aims to enhance the un- derstanding of the distinctive brain patterns associated with ASD. Meenakshi Khosla suggested a volumetric-CNN system that exploits rs-fMRI data’s complete 3D spatial layout. This approach en- ables the creation of non-linear predictive models. Taban Eslami et al. proposed ASD-DiagNet, a framework using fMRI data to classify ASD subjects. They use autoencoder and single-layer percep- tron (SLP) for joint learning to improve feature quality and optimise parameters. Data augmentation through linear interpolation is used to create synthetic datasets for training. Evaluated on a dataset of 1,035 subjects from 17 centers, the model achieves up to 82% accuracy, outperforming other methods by 28%. Xin Yang, Ning Zhang, and Paul Schrader reviewed ASD classification using traditional and DL methods on ABIDE data. Functional connectivity is used for classification using rs-fMRI sam- ples. Three main parts: comparison of brain parcellation techniques, evaluation of connectivity metrics, and comparison of supervised learning models. The findings indicate that the Bootstrap Analysis of Stable Clusters (BASC) holds predictive value for classifying ASD. Additionally, the stability of the correlation metric has been established. Kernel SVM is an optimal classifier with 64.57% sensitivity, 73.61% specificity, and 69.43% accuracy. Ying Chu et al. constructed functional connectivity networks (FCNs) for each subject based on brain region relationships. Traditional studies manually extract network features and build prediction models, requiring expertise. The paper introduces a multi-scale graph representation learning (MGRL) framework that utilizes multiple brain atlases for FCN construction, employs multi-scale-GCNs for feature extraction, and fuses features for ASD diag- nosis. Maya A. Reiter et al. used different ASD sample compositions, considering gender and symptom severity. Random Forest algorithms were employed on resting-state functional connectivity 20 (FC) data from 237 ROIs. Four samples with varied inclusion criteria yielded accuracy rates of 62.5%, 65%, 70%, and 73.75%. Features from cingulo-opercular task control (COTC) network and other net- works were informative. Results suggest classifier performance is influenced by sample composition, yet true ASD heterogeneity challenges utility assessment. Ming Yang et al aimed to distinguish between ASD and TD using brain networks from rs-fMRI. They developed a DL model combining a CNN and attention mechanism, achieving 77.74% accuracy in ASD diagnosis. This suggests that multi-network brain networks could serve as reliable markers for ASD diagnosis, advancing ASD re- search. Wutao Yin et al. developed a three-step approach: constructing brain networks, using an autoencoder to learn advanced features, and training a DNN. The proposed DNN achieved a 76.2% accuracy and 79.7% AUC. Combining DNN with autoencoder reached 79.2% accuracy and 82.4% AUC. The deep learning methods outperformed traditional methods. Amirali Kazeminejad and Roberto C Sotero focused on feature extraction for machine learning, especially considering negative correlations between brain regions. Graph theory-based features are ex- tracted from positive, absolute, and anticorrelation matrices. A MLP is trained via leave-one-site-out cross-validation. Yaya Liu et al. employed an attention-based Extra-Trees algorithm on resting- state functional MRI data from the CC200 atlas. Achieving a mean accuracy of 72.2%, their approach outperforms previous efforts with 2% better precision and 3.2% improved specificity. The model’s con- nectivity analysis reveals pertinent regions for social cognition and interaction, emphasizing reduced correlation in the default mode network in ASD cases. However, the study’s reliance on ABIDE data and its cross-validation strategy might introduce limitations. In the study by Zhi-An Huang et al. , proposed a approach to classification using graphs was presented. They employed DBN alongside the ABIDE database. The process consisted of choosing connectivity attributes through a graph-centered K-nearest neighbors method, which was further enhanced with a path-based algorithm. Automatic hy- perparameter tuning optimizes the DBN. The proposed model outperforms existing methods, achieving a 6.4% improvement on ABIDE. Data augmentation and oversampling identify ASD subtypes. Feng Zhao et al. proposed a high-order D-FCN model based on ”correlation’s correlation” principle, addressing these issues. They extract temporal-invariance properties and use an ensemble classifier to achieve an 83% ASD classification accuracy, identifying unique features at different connectional levels. Michelle Tang et al. presented a deep multimodal model that combines two types of connectomic data from fMRI scans to enhance diagnosis accuracy. The proposed model attains a classification accuracy of 74%, accurately identifies 95% of cases, and achieves an F1 score of 0.805. It surpasses the performance of methods that rely on a single type of data. The study addresses the limitations of using only one functional data type and proposes a more comprehensive diagnostic strategy. Rajat Mani Thomas et al. explores various transformations of rs-fMRI data to prepare it for a 3D convolutional neural network (3D-CNN). Multiple measures are extracted from each voxel’s activity, and regional homogeneity (ReHo) emerges as the best single measure. A 3D-CNN and support vector machine (SVM) achieve competitive results ( 66%) in classifying ASD vs. healthy controls. In their work, Zeinab Sherkatghanad and her team introduced a model that demonstrated a 70.22% ac- curacy when tested on the ABIDE I dataset. This model utilized the CC400 functional parcellation atlas and was built on a CNN architecture. Notably, the model’s computational efficiency suggests it could be valuable for initial assessment of individuals with ASD. Ke Niu et al. , a sophisticated machine learning approach known as the Multichannel Deep Attention Neural Network (DANN) is introduced. This model effectively combines neural networks, attention mechanisms, and feature fusion to process multimodal data. Evaluated on ABIDE repository, with 809 subjects, DANN achieves 0.732 accuracy in ASD classification, surpassing other models. Cross-validation confirms its robustness, offering the potential for future automated ASD diagnosis, countering the time and cost constraints of traditional methods. Amirali Kazeminejad and Roberto C Sotero Utilizing graph theoretical metrics of FC, SVM is employed to extract biomarkers from fMRI data of 816 individuals. The dataset is divided into 5 age groups to consider aging effects. The proposed method outperforms previous studies, attaining impressive levels of precision, sensitivity, and specificity, particularly reaching 95%, 97%, and 95%, respectively. Hailong Li et al. , introduced an innovative approach called the deep transfer learning neural network (DTL-NN) framework. This method integrates data from resting-state functional MRI scans with sophisticated machine-learning techniques. The framework employs a specialized stacked sparse autoencoder to understand the patterns of normal brain connections. This learned information is then applied to classify individuals with ASD. Testing on ABIDE data, DTL-NN outperforms traditional models in accuracy, sensitivity, specificity, and AUC. This suggests DTL-NN’s potential for enhancing neurological disease classification with limited data. Runa Bhaumik et al. employed partial least square regression (PLS) to mitigate site variations in multicenter data. A sparse Multivariate Pattern Analysis (MVVPA) with LOSOCV discriminates ASD from controls across six sites. SVM accuracy (62%) with PLS is comparable to without PLS. Key regions (e.g., Dorsolateral Prefrontal Cortex, Somatosensory Association Cortex) associated with executive function, speech, and perception are highlighted. In the research conducted by Bernas et al. , they investigated the time-based patterns of brain activity using wavelet coherence maps obtained from resting-state networks. They introduced a novel measure called ”time of in-phase coherence,” which was employed to train classification models. Their approach demonstrated an accuracy of 86.7%, 21 with a sensitivity of 91.7% and specificity of 83.3%, effectively distinguishing between individuals with ASD and those without ASD in two distinct datasets. The findings indicate a promising method for identifying ASD based on temporal brain dynamics. Omar Dekhil et al. proposed a machine learning-based system that utilizes power spectral den- sities (PSDs) of brain activation areas, employing a stacked autoencoder and probabilistic support vector machines to achieve around 90% sensitivity, specificity, and accuracy. Mayssa SOUSSIA and Islem Rekik introduced a novel approach for ASD detection by utilizing morphological brain networks (MBNs) derived from conventional T1-weighted MRI data. The method involves quantify- ing shape relationships between cortical regions, both at low and high-order levels. A connectomic manifold learning technique employs diverse kernels to estimate brain connection similarities between ASD and control groups, enabling dimensionality reduction for patient clustering. Maryam Akhavan Aghdam, Arash Sharifi, AND Mir Mohsen Pedram employed deep belief networks (DBN), a powerful deep learning algorithm, to combine fMRI, gray matter, and white matter data from ABIDE I and II datasets. DBN is applied to classify ASDs from typical controls. The approach focuses on high-level features, outperforming prior methods with 65.56% accuracy, favoring early diagnosis. Al- though effective, specificity remains a limitation at 32.96%. Anibal Sólon Heinsfeld et al. employs deep learning methods to detect ASD using brain activation patterns. The study uncovers functional connectivity patterns distinguishing ASD participants and achieves 70% accuracy in identification. Notably, anticorrelation between anterior and posterior brain regions aligns with existing research on ASD-related connectivity disruption. The prevalence of DL methodologies, CNN and neural networks, indicates their effectiveness in cap- turing intricate patterns present in functional and structural MRI data [137, 143, 152]. Furthermore, the recurrent integration of multi-modal data sources underscores the significance of combining various data types to enhance the accuracy and robustness of ASD diagnostic models [149, 153]. The ex- plicit focus on model robustness through cross-validation strategies and data augmentation techniques illustrates the field’s growing concern for developing reliable and generalizable diagnostic tools [139, 147]. The consistent identification of specific brain regions, including the inferior frontal gyrus and the default mode network, highlights the presence of recurring neural markers associated with ASD across multiple studies ,. While many studies report remarkable accuracy rates, the persistent challenge of achieving high specificity rates remains a crucial consideration [143, 159]. The utilization of large datasets, such as the ABIDE, underscores the significance of comprehensive investigations and rigorous cross-validation procedures to ensure the models’ generalizability [147, 150]. While the studies highlighted in the discussion significantly advance ASD detection, they do exhibit certain limitations and research gaps. Notably, many studies heavily rely on specific datasets like ABIDE, potentially introducing biases that could hinder the broader applicability of findings. Additionally, although multi-modal data fusion is emphasized, the optimal techniques for integration remain a continuing research challenge, leading to variations in feature selection and fusion strategies across studies [143, 149]. The potential impact of sample compositions, encompassing variations in gender, symptom severity, and age groups, on model performance, is another limitation [142, 153]. Achieving high specificity in various studies remains an ongoing challenge, necessitating further refine- ment of models to enhance their ability to accurately differentiate ASD from other conditions [152, 159]. Moreover, while DL models often achieve high accuracy, their interpretability poses a research gap, requiring a deeper understanding of the learned features and patterns for effective clinical decision- making [143, 154]. Furthermore, potential overfitting due to limited sample sizes underscores the need for extensive validation on larger and more diverse cohorts to ensure the robustness of the developed models [128, 139]. Researchers have undertaken various efforts to address the issue of data heterogeneity in the ABIDE-I dataset, aiming to identify ASD effectively. In the study by Liu et al. , the dataset consisted of 1102 subjects and 116 Regions of Interest (ROI). They tackled the heterogeneity problem by employing a low rank and class discriminative representation (LRCDR) domain adaptation approach, achieving an accuracy of 73.1%. Similarly, Wang et al. also worked on the ABIDE-I dataset, w