AIMAFE: Autism spectrum disorder identification with multi-atlas deep feature representation and ensemble learning

Highlights

• Multi-atlas functional connectivity is calculated as the original feature representation.

• Multi-atlas deep feature representation is extracted by a deep learning method.

• An ensemble learning strategy is proposed to perform the final ASD identification task.

• Our proposed method achieves classification accuracy of 74.52%.

Abstract

Background

Autism spectrum disorder (ASD) is a neurodevelopmental disorder that could cause problems in social communications. Clinically, diagnosing ASD mainly relies on behavioral criteria while this approach is not objective enough and could cause delayed diagnosis. Since functional magnetic resonance imaging (fMRI) can measure brain activity, it provides data for the study of brain dysfunction disorders and has been widely used in ASD identification. However, satisfactory accuracy for ASD identification has not been achieved.

New method

To improve the performance of ASD identification, we propose an ASD identification method based on multi-atlas deep feature representation and ensemble learning. We first calculate multiple functional connectivity based on different brain atlases from fMRI data of each subject. Then, to get the more discriminative features for ASD identification, we propose a multi-atlas deep feature representation method based on stacked denoising autoencoder (SDA). Finally, we propose multilayer perceptron (MLP) and an ensemble learning method to perform the final ASD identification task.

Results

Our proposed method is evaluated on 949 subjects (including 419 ASDs and 530 typical control (TCs)) from the Autism Brain Imaging Data Exchange (ABIDE) and achieves accuracy of 74.52% (sensitivity of 80.69%, specificity of 66.71%, AUC of 0.8026) for ASD identification.

Comparison with existing methods

Compared with some previously published methods, our proposed method obtains the better performance for ASD identification.

Conclusion

The results suggest that our proposed method is efficient to improve the performance of ASD identification, and is promising for ASD clinical diagnosis.

Introduction

Autism spectrum disorder (ASD) (Amaral et al., 2008) is a neurodevelopmental disorder that affects social interaction and communication (Kim et al., 2011). Families of patients with ASD face a nonnegligible economic burden (Ou et al., 2015). The clinical diagnosis of ASD based on behavioral criteria, assessed using tools such as the autism diagnostic observation schedule (ADOS) (Lord et al., 2000) and autism diagnostic interview-revised (ADI-R) (Lord et al., 1994), has been criticized as lacking objectivity (Timimi et al., 2019) and being likely to lead to misdiagnosis or delayed diagnosis (Wang et al., 2017b). In addition, clinical diagnosis of ASD relies on professional doctors which consumes a lot of medical resources. Thus, it is appealing to develop a more convenient and objective diagnostic method to identify ASD (Yerys and Pennington, 2011).

Currently, magnetic resonance imaging (MRI) has provided a way for clinicians to understand the pathophysiology of brain disorders, such as schizophrenia (Liu et al., 2017, Liu et al., 2018d), ASD (Kong et al., 2019, Wang et al., 2017a) and Alzheimer's disease (Liu et al., 2018b, Liu et al., 2018c). MRI is an imaging technology that can recognize brain disorders based on brain structure. MRI is also a low-cost and non-invasive diagnostic tool which has been widely used in medical field. Specifically, since functional MRI (fMRI) (Huettel et al., 2004, Fox and Raichle, 2007, Buxton, 2009) can infer brain activity by measuring blood oxygen level signals over time, it has attracted increasing attention from researchers engaged in brain dysfunction research (Iidaka, 2015, Dvornek et al., 2017, Liu et al., 2020, Xiang et al., 2020). Brain activity is quantified by changes in the intensity of fMRI images during the acquisition time, which is usually represented by a time series. In general, brain disorders are not just abnormalities in one or several brain regions, but an abnormal connectivity between brain regions. Therefore, to explore the correlation of certain activities between brain regions, functional connectivity (FC) has been proposed and is widely used for the classification and prediction of brain disorders (Du et al., 2018).

In the past decade, some researchers have combined FC and machine learning methods (such as support vector machines (SVMs)) to classify patients with ASD from typical controls (TCs) based on fMRI data (Plitt et al., 2015, Anderson et al., 2011, Chen et al., 2016, Jahedi et al., 2017, Abraham et al., 2017). For example, Chen et al. (2016) first calculated two FCs based on the Dosenbach atlas (Dosenbach et al., 2010) and two frequency bands: slow-4 frequencies (0.027–0.073 Hz) and slow-5 frequencies (0.01–0.027 Hz) which were proposed by Zuo et al. (2010), and then combined the top ranked features of these two FCs via F-scores to perform ASD classification. Jahedi et al. (2017) first extracted FC based on the brain atlas proposed by Power et al. (2011) as the original features of each subject, and then used a conditional random forest (CRF)-based dimension reduction algorithm to obtain more discriminative features to perform ASD classification. Abraham et al. (2017) first extracted a group brain atlas based on a multisubject dictionary learning (MSDL) algorithm, and then calculated the FC based on the group brain atlas using a tangent space embedding algorithm to obtain the features for ASD classification. In recent years, deep learning methods (LeCun et al., 2015) have shown great potential in many areas, including medical image analysis (Litjens et al., 2017, Liu et al., 2018a). Compared with traditional machine learning methods, deep learning methods can automatically learn more hidden features from raw data with multilayer neural networks. For example, Heinsfeld et al. (2018) proposed an ASD identification method that used a stacked denoising autoencoder (SDA) (Vincent et al., 2010) for pretraining to learn a deep feature representation from FC based on the Craddock 200 (CC200) (Craddock et al., 2012) atlas. Parisot et al. (2018) first calculated FC based on the Harvard-Oxford atlas as the original features of each subject, and then used graph convolutional networks (GCNs) to learn the deep feature representation from the FC of each subject to perform ASD classification. Table 1 summarizes some existing FC-based ASD identification methods. To date, although some FC-based studies have obtained relatively good ASD identification results, satisfactory results have not been obtained in clinical practice. Therefore, this is still a challenging problem for ASD identification.

To address this challenging problem, in this study we propose an improved ASD identification method using multi-atlas deep feature representation and ensemble learning which we term “Autism spectrum disorder Identification with Multi-Atlas deep Feature representation and Ensemble learning (AIMAFE)”. First, we calculate three FCs between the time series of brain regions based on three different brain atlases. Then, to obtain a more discriminative feature representation for ASD identification, we propose a multi-atlas deep feature representation method based on an SDA. Finally, we propose a multilayer perceptron (MLP) and an ensemble learning method to combine multiple deep feature representations to perform the final ASD identification task. Our proposed method is evaluated on the Autism Brain Imaging Data Exchange (ABIDE) dataset.