Autism Spectrum Disorder, known as ASD, is a complex neuro-developmental disorder and has been characterized by a series of symptoms including early-onset difficulties in social communication as well as restricted, repetitive behaviors and interests (Pagnozzi et al., 2018). The symptoms of ASD generally occur within the first 3 years of life and tend to last even one's whole life (Hazlett et al., 2017). ASD brings significant impairments on an individual's language, emotions, behavior, self-control, learning, and memory and also is accompanied by intellectual disability. Moreover, it is reported that patients with ASD are far more likely to encounter premature death than healthy controls (Hirvikoski et al., 2015). According to the Morbidity and Mortality Weekly Report (MMWR) Series published by the Centers for Disease Controls and Prevention (CDC) in the United States, the prevalence of ASD among children has increased from 1 in 150 to 1 in 54 over 16 years (from 2000 to 2016) and the incidence rate of ASD was 4.3 times higher in boys than girls (Maenner et al., 2020) in 2016. For each patient with ASD, the average lifetime social cost is approximately $3.6 million (Cakir et al., 2020).

Actually, if ASD is unable to be detected and intervened at an earlier age, the impairments are irreversible. Therefore, early and accurate identification and diagnosis are crucial to improving the life quality of ASD patients and their families. Unfortunately, it is notoriously difficult to diagnose, especially in children, since the cause of ASD is a result of combined factors, including genetics, the structure and function of the brain, as well as environmental influences (Rakić et al., 2020). Until now, there are still no effective medical treatments for ASD. For the current practice guidelines to assess, diagnose and treat ASD, it is recommended to use the behavioral observation of symptomology following the Diagnostic and Statistical Manual (Fifth Edition) (DSM-5) (American Psychiatric Association, 2013) symptom criteria and the International Classification of Mental and Behavioral Disorders (Tenth Edition) (ICD-10) (Organization, 1993). However, uniformity is lacking while using these practice guidelines, so it is probably prone to misdiagnosis (Eslami et al., 2021). Furthermore, these guidelines cannot point out the biological bases related to behavioral symptoms due to unclear neuroanatomy. Finally, these limitations have resulted in calls for more optimal diagnostic approaches for ASD.

In the last few decades, advances in non-invasive neuroimaging techniques and analysis have provided crucial knowledge to uncover patterns of brain structure and function that would be symptomatic for the autism spectrum. The vast majority of statistical methods on Structural MRI have intended to explore the common patterns between patients with ASD and healthy groups, but previous volumetric and morphological analysis on structural MRI often has derived contradicted results. For example, some research work reported decreased volumes of the amygdala for ASD (van Rooij Daan et al., 2017) while others did not find significant alterations (Maier et al., 2015). Focusing on the hippocampus volumes, some reported its reduction, others reported its enlargement or no changes (Barnea-Goraly et al., 2014; Maier et al., 2015). Xiao et al. (2014) has found that both gray and white matter (WM) has a significant increment with ASD, and Hazlett et al. (2017) has pointed out brain volume overgrowth is related to the emergence and severity of ASD. While Palmen et al. (2005) and Jou et al. (2011) have noted that there is no difference or decreased WM volume between ASD and healthy controls, and Riddle et al. (2016) conducted voxel-based morphometry analysis and found that the total brain volume and the left anterior superior temporal gyrus increased for children aged 2–4 with ASD. But these brain structural abnormalities are subtle at later ages (Riedel et al., 2014). These inconsistent findings are most likely due to different collecting approaches and limited sample size with heterogeneous characteristics of subjects (Riddle et al., 2016). Moreover, traditional statistical analysis is based on mass univariate techniques which process a single voxel independently and ignore the relationship between voxels (Bonnici et al., 2012; Samartsidis et al., 2016). Furthermore, it defines the common pattern at the level of groups and is unable to predict the unknown sample at the level of individuals (Zhutovsky et al., 2019; Hu et al., 2021).

Most recently, the rapid advance of machine learning has made it becomes possible to explore the underlying neural mechanisms and provide accurate predictions and convincing explanations for ASD from various aspects (Khodatars et al., 2020; Eslami et al., 2021). Knutson (2013) has pointed out that machine learning can detect differences in neuroimaging data that might not be detected with traditional univariate analysis. In previous studies, typical statistical machine learning and deep learning have been utilized to identify ASD from NC in terms of structural and functional alterations. Statistical machine learning requires the design of handmade features (feature extraction/feature selection) and implement the identification of patients with ASD based on these features (feature classification). Ecker et al. (2010) has applied SVM to investigate the whole-brain differences of GM and WM volume on 44 subjects and obtained significant predictive power. Additionally, it has been found that these brain differences are related to symptom severity. Ecker et al. (2010) and Wee et al. (2014) have extracted morphological features based on structural images and used SVM or multi-kernel technique to achieve satisfactory results. Furthermore, Zheng et al. (2018) have constructed a multi-feature-based network based on morphological features to explore the cortico-cortical similarities of ASD. Bilgen et al. (2020) have modeled the morphological relationship between pairs of ROIs with a cortical networks and verified the classification performance of different machine learning methods. Concerning female children, Calderoni et al. (2012) have detected the abnormality of the gray matter volume based on SVM-RFE (Leung et al., 2006; ChenZhiHong et al., 2020) and found the increased cortical volume in some brain regions involving the left superior frontal gyrus (SFG). In addition, bilateral SFG and right temporoparietal junction (TPJ) resulted in the appearance of some atypical symptoms of ASD and might be relevant to the pathophysiology of female children in ASD. These findings are helpful to reveal the important influence of the structural alterations and the relationship between the brain structure and the pathophysiology of ASD.

However, the lack of a sufficiently large sample at a single site probably leads to poor generalizability that is notably serious for neuroimaging due to limited participants and super-high dimensionality of data. Consequently, the investigation of large sample data from multi-site has attracted increasing attention. Some studies (Spera et al., 2019; Mwiza et al., 2020) have figured out the superiority of machine learning for the classification of multi-site data based on fMRI or the combining structural and MRI in the Autism Imaging Data Exchange database (Di Martino et al., 2017). Due to the excellent performance of deep learning in the field of artificial intelligence on large sample data, some researchers have begun to detect abnormalities of functional connectivity based on deep learning. Deep learning has combined dimensionality reduction and feature classification and implemented the end-to-end classification model automatically, which has achieved satisfaction performance (Eslami et al., 2019; Sherkatghanad et al., 2020). Furthermore, some attempts have been done to fuse structural and functional features with the model of deep learning to improve the classification performance (Rakić et al., 2020). But it cannot be denied that deep learning handles data with the mechanism of a black box and it is so hard to identify the abnormal brain regions and connect the classification accuracy with the underlying mechanism of ASD. Furthermore, multi-site data also has brought the issue of data heterogeneity due to different scanning parameters and participant populations. The direct way to address the heterogeneity issue is to apply dimensionality reduction to transform source data into features in the field of machine learning (Wang et al., 2020). Furthermore, these studies also utilized leave-one-site-out cross-validation to evaluate the classification performance in the expectation of reducing the impact of heterogeneity simultaneously (Rakić et al., 2020; Eslami et al., 2021). In order to make the results robust, Ashourvan et al. (2016) have further proposed intra-site cross-validation and inter-site cross-validation and achieved 65% accuracy with functional connections (FC) to identify ASD from normal control.

In fact, detecting the structural/functional brain alterations is vital to reveal the pathological mechanism of ASD. In particular, these brain regions with obvious differences can be recognized as the neuro-imaging biomarkers related to the disease. Based on this kind of neuro-imaging biomarkers (brain regions), achieving excellent classification performance even from different multi-sites would be the most desirable and helpful in the clinical diagnosis. Meanwhile, aiming at a few investigations on volumetric changes based on machine learning, this study has applied machine learning techniques followed the unified framework to implement model-level, feature-level, and biology-level assessment successively. First of all, a searchlight-based classification method has been used to detect the volumetric changes locally and some candidate brain regions have been defined based on the areas of the volumetric changes at the model-level assessment; Regarding distinguished brain regions, this study has processed the “visual lesion” analysis at the feature-level assessment. Stability based on nested cross-validation and multi-site validation of each region has been evaluated. The candidate regions with good stability performance have been preserved and considered as candidate biomarkers related to ASD. Finally, this study investigated the relationship between candidate biomarkers and symptom severity and analyzed our results with previous findings.

The main contributions of the study are discussed as follows: (1) Previous machine learning studies on ASD mainly focus on the classification performance or the important features. Furthermore, this study paid attention to the interpretability of the machine learning model to explore abnormal brain regions related to ASD and conducted model level and feature level assessment to ensure the robustness and stability of the results. (2) The correlation analysis between abnormal brain regions and clinical severity in our study has further proved the relationship between the volume changes of some specific brain regions and the clinical symptom. (3) The findings in our study are partly consistent with previous research work. The abnormal gray matter (GM) volume in the temporal lobe, Broca and Wernicke area probably provides the support for the social brain hypothesis and the broken mirror theory of ASD, which is helpful to understand the neuroanatomy of ASD.

The structure of this study is as follows: First, in section 2, we provide a brief introduction to the pre-processing procedure and statistical analysis of sMRI data. In section 3, we describe the machine learning workflow in detail. Experimental results and discussion are provided in section 4. Finally, in section 4, we conclude the study and discuss the future direction.

This study aimed to identify the brain abnormality and predict ASD from TDC via machine learning techniques. However, neuroimaging-based ML models like the “black-box” and unable to be understood from the prospect of neuroscience. To address this issue, Kohoutov et al. (2020) has developed a unified framework to enhance the interpretability of ML models and provide mechanistic insights into underlying neural or disease processes. The proposed framework contains a three-stage process of assessment including Model-level assessment, Feature-level assessment, and Biology-level assessment. In the first stage, the ML model has been built from observations and assessed in terms of its sensitivity, specificity, and generalizability. In the second stage, significant features have been identified from a prediction within the model. Finally, the neuroscientific plausibility of the ML model has been proved with evidence from previous literature and other studies.

However, ML models based on neuroimaging are often built on numerous features and limited participants, which makes the model is prone to overfitting and leads to poor generalization and expensive computational cost even if dimensional reduction techniques have been used. Moreover, isolated features are often insufficient to acquire satisfactory predictive performance and explain the model performance. Consequently, the study has designed a neighborhood-to-regional machine learning workflow within this unified framework to identify structural alteration and discriminant ASD from TDC. The workflow proposed in the study has been illustrated in Figure 1.

Current VBM findings have delineated brain regions with consistently increased or reduced GM volume (Cauda et al., 2011). In this study, statistical analysis has reported a widespread reduction of GM volume in ASD with 7-13 years old. Research on brain development in ASD across the lifespan has demonstrated a complex neurodevelopmental trajectory, characterized by an early brain overgrowth followed by undergrowth in middle childhood and early adolescence (Courchesne et al., 2001). This might support the findings of statistical analysis in our study.

Our results based on machine learning have demonstrated that a widespread structural alteration of GM volume involved in bilateral superior temporal gyrus, bilateral middle temporal gyrus, left inferior temporal gyrus, right orbital SFG, bilateral opercular inferior frontal gyrus, left orbital inferior frontal gyrus, right rolandic operculum, right olfactory cortex, right gyrus rectus, right insula, right inferior parietal lobe with Supramarginal gyrus and Angular gyrus. Especially, multi-site dataset validation also has verified the robustness of the machine learning framework with three-level assessment. Since ASD is a complex neurodevelopmental disorder, involving language, reading, emotion, social interaction impairments, the quantitative meta-analysis in Geschwind and Levitt (2007) and Maximo et al. (2014) have suggested that ASD is unlikely to be associated with the abnormalities in one specific region alone but to be linked to the abnormalities of multiple, spatially distributed, neural systems. The finding may shed light on the widespread differences in GM volume found in our study.

The findings in the study have almost covered the whole temporal lobe including bilateral superior temporal gyrus, bilateral middle temporal gyrus, and left inferior temporal gyrus. Since attention has been directed to explore the neurobiological mechanism of ASD first, the abnormality of the temporal lobe has been speculated to link with the deficits in language and social behavior of patients with ASD (Hauser et al., 1975; Bachevalier, 1994; Kates et al., 2010). Ritvo et al. (1986) has examined the brains of four autistic subjects and found the localized pathological changes in the temporal lobe from autopsy-based research. It is considered that the superior temporal gyrus is a potential import biomarker of ASD (Pierce, 2011; Sophia et al., 2013). Based on the VBM-Dartel technique, Riddle et al. (2016) have revealed enlargement of the left anterior superior temporal gyrus in ASD. It is believed (Bigler et al., 2007) that the superior temporal gyrus plays a crucial role in social cognition, which participates in auditory and language processing. On the other hand, the VBM analysis has found reduced GM volume in the middle temporal gyrus (Kohoutov et al., 2020). According to analyze structural images of low functioning ASD children from 2 to 10 years old, the reduction of GM volume in the left inferior temporal gyrus has been identified, which appears to be involved in visual object perception (Riva et al., 2013). Similarly, RT et al. (2000) have revealed the fMRI (functional MRI) alterations of the inferior temporal gyrus when engaging in facial recognition tasks. Zilbovicius et al. (2000) has also found the localized dysfunction of the temporal lobe from the aspect of PET. Brothers (2002) has proposed the concept of the social brain first in 1990, which was defined as a group of interrelated neuroanatomical structures which are used to process social information, recognize other individuals and evaluate their psychological state, including intentions, dispositions, desires, and beliefs. The temporal lobe plays a very important role in the hypothesis of the social brain. The posterior superior temporal sulcus recognizes biological movements, such as eyes, hands, and other body movements and helps to interpret and predict others' behavior and intentions (Allison et al., 2000). The fMRI study of patients with ASD has shown that the differences in the activation on temporal lobe compared with their families and normal people, and the worse the social ability, the weaker the activation. Furthermore, it has been found that the degree of activation was positively correlated with the clinical manifestations of social impairment (Sugrue et al., 2010). The abnormality of the temporal lobe found in this study not only supports the hypothesis of the social brain but also suggests that the area of temporal lobe abnormalities in patients with autism may be larger than that found in previous studies. However, the hypothesis still needs to be further confirmed by quantitative and qualitative autopsy reports and animal studies.

Furthermore, the study has found GM abnormalities in the Broca area (posterior frontal lobe corresponding to BA44) and Wernicke area (superior marginal gyrus corresponding to BA39 and angular gyrus corresponding to BA40). Meanwhile, we also found negative correlations between GM volume in the right angular gyrus and verbal/nonverbal communication score in ADI_R. Language deficits are the core diagnostic characteristics in ASD and both of Broca area and Wernicke area are associated with language understanding. Adam et al. (2004) found that the Wernicke area, which is responsible for the understanding of single words, is more active than the Broca area, which is responsible for the understanding of complex sentences and has proposed the “underconnectivity theory” to explain why some patients with ASD have excellent ability to process single words, rather than complex sentences. Osbarn (2020) has found weakened functional connectivity in the area of Wernicke. Recently, researchers have established “the broken mirror theory” of autistic patients. It is supported that the dysfunction of the Human Mirror Neuron System (MNS) is the main cause of social and cognitive deficits in ASD (Vivanti and Rogers, 2014). The Broca area in humans has been considered as homologous to F5 as a part of MNS. The results in our studies also supported the MNS dysfunction in ASD individuals. However, the relevant evidence about the role of MNS in ASD still is not enough which urges us to build a more perfect MNS theory to understand the causes of social communication disorder in ASD (Southgate and Hamilton, 2008).

For other regions identified in the present study, they have also been reported in previous literature. Shijun (2021) has constructed a three-dimensional residual network based on deep learning and found the GM reduction in the orbital inferior frontal gyrus and Rolandic operculum. Riva et al. (2013) also found the reduced GM volume in the orbital part of the inferior frontal gyrus. In light of the meta-analysis based on large samples, the volume of GM in the insula and inferior parietal lobe decreased (Cauda et al., 2011). Li et al. (2019) have found that the GM volume of the bilateral gyrus rectus decreased, and the left rectus was negatively correlated with the clinical symptom score. Although it is not reported that the volume changes in the olfactory cortex of patients with ASD, the olfactory cortex is located at the anterior bottom of the limbic system and reciprocally connected with other structures, such as the amygdala, hippocampus, hypothalamus, the olfactory cortex is related to emotion and memory. As a consequence, the abnormal GM in the olfactory cortex in ASD may lead to emotional and memory problems. It is worth mentioning that some current studies on ASD have found GM differences in the cerebellum, hippocampus, and parahippocampal gyrus (Faridi and Khosrowabadi, 2017; Lotze et al., 2019). In particular, a decreased number of Purkinje cells in the cerebellar hemisphere have been observed (Ritvo et al., 1986). The study only found that the differences of the parahippocampal gyrus during statistical analysis. For model-level assessment, we have detected the alterations in these three brain regions, including the bilateral hippocampus, and parahippocampal gyrus and left crus in the cerebellum. However, they have been excluded from the candidate brain regions due to the poor performance in stability analysis. Traut et al. (2018) have compared and analyzed cerebellar volume of a large sample of ASD patients with normal subjects, and reported that the change of the cerebellar volume was significantly correlated with age, gender, and IQ rather than ASD diagnosis. Even though some studies have reported the differences of WM in ASD (Ecker et al., 2010; Xiao et al., 2014; Górriz et al., 2019), our study has not found abnormal white matter based on structural images. At present, contradictory conclusions often have been derived from a variety of ASD research due to the heterogeneity of subjects, including different subtypes, different scanning parameters from different centers, different ages and genders in ASD (Pua et al., 2017; Hiremath et al., 2021). In addition, the inconsistent findings of research work based on machine learning might be induced by different feature exaction techniques. For example, Haar et al. (2016) have achieved poor classification based on morphological features of ROI on the multi-site dataset and suggested that anatomical abnormalities may be only present in some distinct subgroups of ASD, while Zheng et al. (2018) have obtained superior classification performance with multi-feature-based networks based on morphological features.

Our results suggest that structural MRI can provide neuroimaging-based biomarkers for ASD. Such biomarkers could be used to complete and improve the diagnosis and treatment of ASD clinically (Walsh et al., 2011). On the one hand, we can utilize the classification performance of these identified biomarkers based on the machine learning model proposed in the study to improve diagnostic accuracy. On the other hand, the behavioral social malfunctioning in ASD might be modified by neural or behavioral treatments. For example, it is also reported that the behavioral training of facial expression communication behavior can help to improve the neural activities of ASD patients related to some social brain regions, such as MTG, so as to improve their capability of expression recognition (Bölte et al., 2015).

Although our results are consistent with some previous reports, several limitations of our study should be acknowledged. First, the confounding factors should be considered. ASD is a complex disease with multiple confounding factors, such as age, gender, IQ, and the inherent heterogeneity of the disorder. Our studies attempted to control age and gender in statistical analysis but failed to find an appropriate approach to remove the influence of the confounding factors in machine learning methods. On the other hand, the controlling of confounding factors is still highly controversial in the studies of ASD (Thomaidis et al., 2015). Some researchers have claimed that IQ should be strictly matched or statistically regressed out, while others have argued that the variability truly associated with ASD could be also discarded as “non-specific” when attempting to control some non-specific factors, such as IQ (Osbarn, 2020). Some studies have investigated the gender differences in ASD (Halladay et al., 2015; Prosperi et al., 2021) due to a high incidence rate of ASD in boys. Second, the current study took advantage of the relatively large sample of participants with ASD in the ABIDE II database to process multi-site validation via machine learning. However, we have to point that all three databases used in our study were collected by Siemens scanner with similar scanning parameters, which might make a less dispersion among data. In the future, the investigation of ASD about neuroanatomical alterations on larger samples from diverse clinical and demographic subgroups will significantly promote understanding neuropathology mechanism in ASD.

In this study, the VBM analysis has revealed a widespread reduction of GM volume when comparing ASD with TDC. Furthermore, our machine learning analysis followed the unified machine learning framework has revealed candidate neuroimaging biomarkers related to ASD and confirmed the relationship between regional GM volume and symptom severity. Our results have suggested that candidate neuroimaging biomarkers are useful to characterize the profile of brain anatomy in ASD and improve the diagnosis performance in clinical applications.

Publicly available datasets were analyzed in this study. This data can be found here: http://fcon_1000.projects.nitrc.org/indi/abide/abide_II.html.

YD, WZ, and CLi conceived and designed this study. HJ, YT, and XL participated in the analysis of MRI dataset. CLu and DY helped to improve the manuscript. All authors contributed to the article, read, and approved the final manuscript.

This work is supported by the China Postdoctoral Science Foundation (No. 2016M592656), Sichuan Science and Technology Program (No. 2018JY0272), Science & Technology Bureau of Chengdu (No. 2020-YF09-00005-SN), Erasmus+ SHYFTE Project (No. 598649-EPP-1-2018-1-FR-EPPKA2-CBHE-JP), and the Key Laboratory of Pattern Recognition and Intelligent Information Processing, Institutions of Higher Education of Sichuan Province (No. MSSB-2021-12).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.