Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by difficulty in social communication, repetitive patterns of behavior, and cognitive deficits [1]. Depending upon the severity of the condition, individuals may require a minimal to a substantial amount of support to carry out even daily activities. The underlying cause of ASD is still under investigation, but the existing studies point out that genetic and environmental factors may be responsible for this condition [2]. Current findings have shown that the estimated prevalence of ASD has been rising over time on a global scale and also for various subpopulations [3]. ASD diagnosis is challenging due to the lack of validated clinical diagnostic tests and heavy reliance on subjective evaluations and behavioral data like developmental history and cognitive test scores [4]. Such methods are prone to error and require highly skilled examiners to make a diagnosis based on the described symptoms. Because of the lack of objectivity in the condition description, most diagnoses are prolonged, in addition to being time-consuming [5]. Thus, there is a need for an automated diagnostic approach to detect ASD based on certain brain anatomical alterations.
Several studies have found morphological changes in the brain that can act as an effective biomarker in the diagnosis of ASD [6, 7]. Multiple modalities like electroencephalography, structural magnetic resonance imaging (sMRI), diffusion tensor imaging, and functional MRI have been used to study ASD [8, 9]. sMRI is a powerful technique that generates accurate, high-quality three-dimensional images of the brain regions. It can recognize the neuroanatomical heterogeneity involved in ASD that can aid in the early detection and treatment of the condition [10]. In our study, we have used the sMRI data to analyze the morphological changes of the ASD brain.
The morphological changes observed in the brain include abnormalities in the sizes of different cerebral, cerebellar, and subcortical regions in individuals with ASD [11–16]. Variations in microstructural integrity, and grey and white matter volume have also been observed in various investigations and could have an impact on interhemispheric connectivity [17, 18]. Researchers have used various measures like changes in volume, thickness, surface area, local gyrification index, and mean curvature of the cortical and subcortical regions of the brain to identify these morphological variations from the sMRI [19–21]. Attempts have also been made in the past to map individual brain networks using voxel-based morphometry measures and wavelet transform to establish morphological connectivity [22]. A study conducted by Mensen et. al. [23] emphasized the significance of differentiating between cortical surface area and thickness in studying cortical development, and it was implied that the development of the cortical surface area is relevant for ASD. In our study, we used the surface area as the morphological feature (MF) and compared its performance with the morphological distance-related features (MDRF).
The use of machine learning methods has been widely employed for the classification of ASD using sMRI features [24]. Parametric classifiers such as logistic regression [25], non-parametric such as support vector machine (SVM) [26], random forest (RF) [27], and neural network [28] have been used to diagnose ASD and typical development (TD). However, not many of them have focused on the comparative evaluation of the correlation between brain morphometric features. Ecker et. al. [26] investigated the predictive power of the whole brain MR scans and found out that SVM can identify even small and widely dispersed changes in the brain networks between adult ASD brain and TD. Xiao et.al. [27] used features like regional cortical thickness, cortical volume, and cortical surface area and tested the efficacy of the top 10, 20, and 30 features by feeding them into the RF classifier, reporting it to be an optimal classifier for the smaller sample size. A hybridized model optimizing multilayer perceptron (MLP) reported very high classification accuracy for child, adolescent and adult ASD subjects [29]. In our study, we used RF, SVM, and MLP to evaluate the performance of MF and MDRF.
The first objective of our study was to find the effect of MF and MDRF on the diagnostic classification of ASD. Next, we focused on identifying the various brain regions responsible for the diagnosis followed by finding the optimal machine learning classifier. The paper is organized as follows: In section 2 we described the database used in our study, the parcellation process, feature extraction, feature selection and classification methods. In section 3 we summarized the main results of MF and MDRF methods and their site-specific classification performance. In section 4 we provided a discussion of the results followed by section 5 where we concluded our findings.