Autism Spectrum Disorder Detection Using Parallel Deep Convolution Neural Network and Generative Adversarial Networks

According to the World Health Organization, ASD affects one child in 160, with a probable incidence of anxiety, attention, mental stress, and depression deficit hyperactivity disorder [1]. Mostly, children from 6 to 17 years of age have ASD, a neurological condition that impairs social interaction and communication abilities. The term ASD defines a group of neurological and mental ability development diseases. A few initial symptoms of ASD include abnormal social interaction and emotional control, limited interest, redundant behaviors, and hypo-reactivity or hyper-reactivity to sensory stimuli [2, 3].

Many autistic people face challenges in learning, growing, controlling, interacting, or specific basic living skills. ASD places a significant financial burden on society and the families of sufferers. An early and precise diagnostic framework is required to separate ASD subjects from normal controls. Non-invasive and neuroimaging approaches have recently attracted researchers to contribute to the additional diagnosis of ASD. For the detection of ASDs, a variety of structural and functional brain imaging modalities are frequently utilized, including structural MRI (sMRI) [4], fMRI [5], diffusion MRI [6], electroencephalography (EEG) [7], magnetoencephalography (MEG) [8], and electrocardiography (ECG) [9]. Structural MRI explores the structural characteristics of the brain. Early age ASD identification is crucial, but it will help children with ASD improve their communication skills and social awareness, thus improving their quality of life [10, 11]. An early diagnosis plays an important role in illness management and proper treatment. Creating a model based on interactions between functional or structural brain regions is crucial in diagnosing neurological illnesses, including epilepsy, Alzheimer's, and autism [12-14]. The fMRI is employed to examine the brain and its structures, which recognizes linked changes in the blood oxygen level-dependent (BOLD) signals from the various parts of the brain. The functional connection between different brain areas that affect global brain networks has been demonstrated to be disrupted by ASD.

Figure 1 shows the MRI images of an average person and autism patient. For both the autism and normal groups, there was sizable bilateral activity in the amygdala in the familiar face versus fixation condition. The stranger versus fixation comparison and the stranger versus typical face comparison revealed no discernible amygdala activity in either group.

1.png

Figure 1. MRI images of a normal person and an autism patient [15]

As a result, the primary objective of the research is to categorize individuals with ASD and healthy controls using brain patterns of functional connectivity [16]. The early detection of ASD can help to provide the proper medical and psychological treatment to the patients. The traditional ASD detection schemes include behavioral analysis and monitoring, which is tedious, time-consuming, and depends upon the expert's knowledge. The vast growth in the population and ASD screening has led to an automatic ASD detection scheme that is faster, more efficient, and more reliable [17].

In categorization and representation learning, the usage of CNN has garnered much interest recently [18]. CNNs are robust classifiers performing well in various applications with significant free parameters [19]. Additionally, CNN models can handle many free parameters and have improved feature extraction accuracy. The CNN model consists of the convolutional, fully connected, normalization, and pooling layers. The CNN model can analyze brain biomarkers in individuals with ASD using fMRI. The CNN models better represent spatial and spectral features from the MRI images, which help to characterize the minor variations in the brain region due to autism. CNN can provide a better balance between the local and global features of the MRI images. However, the effectiveness of the CNN is limited due to extensive hyper-parameter tuning, more significant trainable parameters, data scarcity problems, poor feature discrimination, etc. [20, 21].

The limitations of current methods, such as extensive hyperparameter tuning and data scarcity problems, present the need for a more efficient and accurate ASD detection method. This paper proposes a novel methodology using parallel deep learning networks (PDCNN) to address these challenges.

Figure 2 explores the process flow of the suggested ASD detection scheme that includes data augmentation, training, and testing phases to classify normal and ASD patients. The significant offerings of the suggested ASD scheme are summarized as follows:

• Development of novel PDCNN for ASD detection to improve the feature distinctiveness of the fMRI images.
• Implementing a GAN for data augmentation to minimize the data scarcity problem due to varying dataset sizes.

The arrangement of the remaining article is as follows: Section II describes information regarding the recent trends in ASD using deep learning. Section III gives a detailed description of the dataset and data augmentation using the GAN and PDCNN framework. Further, section IV delivers the discussions on simulation outcomes. Lastly, it offers the concise findings of the work and provides future scope.

2.png

Figure 2. Process flow of the proposed method

3.1 Dataset: ABIDE-I

The ABIDE, a collaborative project including neuroimaging and phenotypic data acquired from 1,112 individuals, is the most popular data-driven technique for diagnosing autism and exploring its biomarkers. ABIDE-I dataset is a global multisite database comprising 1,112 samples of phenotypic data and structural, resting-state fMRI from 16 sites, including 539 people with ASD and 573 others [31]. The proposed system uses preprocessed fMRI images from the ABIDE-I dataset. The images are resized to 128×128 pixels for computational simplicity.

3.2 Data augmentation using GAN

The availability of ASD MRI images is challenging due to inadequate resources, unavailability of experts for assessing MRI images, ethical issues, etc. The lower and uneven samples in the training dataset lead to data scarcity and often result in poor ASD accuracy. Thus, data augmentation is essential to create the synthetic dataset for ASD detection. The following Figure 3 shows the GAN network.

3.png

Figure 3. GAN generator and discriminator working [29]

The GAN creates the synthetic data samples from the available samples and increases the dataset's number of samples. It consists of a generator (GN) and discriminator network (DN). The GAN is an unsupervised deep learning architecture. The GN generates the enhanced sample by accepting the arbitrary input. The DN compares the original sample to an enhanced sample produced by the GN to distinguish an authentic sample from a fraudulent one. The original fMRI images are utilized as the input to the GAN network in the deliberate effort to expand the database. The GAN is trained for 500 epochs with an initial learning rate of 0.01 and batch size of 128. The predicted absolute error E(GN, DN) for the GN and DN is given by Eq. (1).

$E(G N, D N)=\frac{1}{2} E_{x \sim P t}[1-D N(x)]+\frac{1}{2} E_{x \sim P Z}[D N(G N(z))]$            (1)

During the training of GAN, GN tries to lessen the error whereas DN aims to increase the error in training and synthetic samples using Eq. (2).

$\min _{G N}\left(\max _{D N} E(G N, D N)\right)$           (2)

Table 2. Original and augmented dataset (ABIDE-I –fMRI)

GN(z) signifies the GN output, z denotes the random noise factor, DN(x) represents the DN output, x represents the training sample, Pt denotes the likelihood of real MRI data, and Pz represents the likelihood of artificial MRI. A total of 100 synthetic images are created per class for every site using GAN to overcome the data scarcity, as described in Table 2. The Adam optimization algorithm trains the GAN algorithm for 200 epochs. The cross-entropy function is used for the loss evaluation during training, and the initial learning rate is set to 0.001.

3.3 PDCNN architecture

The sequential DCNN uses cascaded layers of the CNN layers with uniform convolution filter size. However, different filter sizes can help capture the fMRI image's various local and global characteristics, which can be missed in sequential DCNN. Also, increasing the layer in sequential DCNN may lead to a vanishing gradient problem and a poor detection rate [32-35].

Thus, the proposed PDCNN includes 3 parallel arms of a sequential DCNN with different filter sizes to combine the attributes captured by various filters. Each parallel arm encompasses three layered DCNNs that include the convolution layer (Conv), rectified linear unit (ReLU), batch normalization (BN) layer, and maximum pooling layer (MaxPool) as illustrated in Figure 2. The first, second, and third parallel arm of the PDCNN uses a convolution filter with a size of 3×3, 5×5, and 7×7, respectively. However, every sequential DCNN at the parallel arm uses 32, 64, and 128 convolution filters in each CNN layer.

The original image with 128×128 pixels dimensions is fed to three parallel arms. The output of the third MaxPool layer of all three parallel arms is assembled in vector, which is further provided to a fully connected layer (FC) that connects each neuron of the parallel and sequential arm and help in boosting the connectivity and correlation between different global and local features of the fMRI image. Lastly, the Softmax classifier layer (Softmax) is employed to classify regular and ASD patients.

3.4 Deep convolutional neural network

The CNN has shown good spatial representation capability for biomedical images. It offers high-order abstract-level features and assists in capturing hierarchical features. Each DCNN layer includes Conv, BN, ReLU, and MaxPool layers. Increasing the Conv layers boosts the feature distinctiveness of the fMRIs.

A. Convolution Layer

In the Conv layer, the fMRI is converse with a filter size of w × w, as given in Eqs. (3) and (4). It transforms the input brain fMRI into high-dimensional feature maps. Because of its high level of abstraction, this layer is frequently known as a hidden feature extractor.

$c(x, y)=i m * f$             (3)

$C(x, y)=\sum_{i=1}^{r o w} \sum_{j=1}^{c o l} i m(x-i, y-j) . f(i, j)$,              (4)

where, im denotes the original picture, C denotes the Conv layer output, and f  represents the Conv filter kernel. Eq. (5) can determine the Conv layer output's dimensions while accounting for padding and striding.

$D_{\text {out }}=\left[\frac{D_{\text {in }}-2 P-w}{s}+1\right]$           (5)

Here, the filtered measurements are w. The parameters p and s are padding sizes in pixels, striding value in pixels, and initial and output image dimensions, respectively.

Eq. (6) provides the Conv layer output's multidimensional dimensions.

$[r o w, c o l, k] *\left[w, w, N_f\right]=\left\{\frac{r o w+2 p-w}{s}+1, \frac{c o l+2 p-w}{s}+1, N_f\right\}$,            (6)

where, $N_f$ indicates the number of image channels, w depicts the dimension of the filter, row represents the fMRI’s width, col denotes the image's height, $N_f$ denotes the number of Conv filters, and S means the stride value. Before training, the Conv filter weights are set randomly and then changed to their optimal value using learning algorithms like the Adam optimizer, stochastic gradient descent (SGD), RMSProp, and mini-batch gradient descent (MBGD).

ASD is a severe disease in the later stages, giving rise to high depression and anxiety and may lead to severe health problems. Hence, its detection is to be done in the early stages. Researchers have proposed different methodologies for its detection. Many researchers for the detection of ASD present traditional CNN. The dataset samples in other publicly available databases for experimentation are of limited sizes. There are certain constraints to employing CNN, such as extensive hyper-parameter tuning, data scarcity, etc. Limited dataset samples could give rise to overfitting during the training of neural networks. Hence, it is necessary to enhance the sample size of the dataset.

Thus, a novel methodology is proposed and implemented in this work that utilizes PDCNN for detection purposes and GAN for data augmentation to increase sample sizes of datasets.

The PDCNN assists in acquiring a better correlation between the connectivity and correlation features of fMRI images for ASD detection with the help of a parallel arm with variable filter size. The proposed PDCNN achieved remarkable ASD detection accuracy with fewer trainable parameters than the traditional state-of-the-art.

The accuracies of the models proposed by different researchers range from 62% to 88%, with trainable parameters not less than 1268160. The proposed PDCNN provides improved accuracy of 90.63% and 920130 trainable parameters, significantly improving over traditional state-of-the-art CNN methodologies. The proposed PDCNN is less complex, lightweight, and faster than the previous state of the arts. The GAN-based data augmentation helps to tackle the data scarcity difficulty arising from lower and unequal training samples in the data sets.

In the future, the results of the PDCNN could be improved for massive data to increase ASD detection accuracy. Also, the system's performance could be analyzed in real scenarios by collecting actual dataset samples to generalize the proposed methodology. The effectiveness of the PDCNN can be studied for various genders and age groups. The PDCNN method can further be used to classify the grades of detected ASD and explore the risk and severity of the disease.