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

Autism spectrum disorder (ASD) is becoming a crucial issue in ages 6 to 17. This disease causes a neurological state that affects social interactions and communication abilities. ASD introduces depression, anxiety, hyperacidity, etc. It may lead to severe disorders in the patients. So, its diagnosis is essential in the early stages. Brain MRI is the popular diagnostic tool used for the detection of ASD. Moreover, with technological advancements, many sophisticated and proven techniques need to be developed for ASD detection. Advanced machine learning and Deep Convolution Neural Networks (DCNN) have attracted the attention of researchers for various applications such as image classification, automotive software engineering, and speech recognition, enabling significant progress in neuroscience. The DL supports improved computational intricacy, the ability to handle larger data, and the high efficiency of the algorithm. However, the DCNN is a well-known algorithm most commonly used for neuro-imaging applications due to the requirement of extensive hyperparameter tuning, data scarcity problems, and inadequate feature representation. This paper discusses ASD detection with functional magnetic resonance imaging (fMRI) using parallel DCNN (PDCNN). The PDCNN helps to acquire distinctive features with different filter kernels at parallel layers to describe the distinct local connectivity features of fMRI images and improve ASD detection accuracy. Also, a Generative Adversarial Network (GAN) is employed for data augmentation, which helps to generate synthetic realistic MRI samples by learning the fundamental distribution of the inputs to diminish the data imbalance problem. The performance of the proposed system is evaluated with a multisite dataset named the Autism Brain Imaging Exchange (ABIDE-I). The suggested PDCNN gives an accuracy of 90.63%, precision of 0.96, recall of 0.87, and F1-score of 0.92. The suggested PDCNN provides improved results and utilizes fewer trainable parameters than the traditional methods.