Enhancing Autism Spectrum Disorder Diagnosis through a Novel 1D CNN-Based Deep Learning Classifier

Autism spectrum disorder (ASD) is a neurodevelopmental state that becomes evident during early childhood. It is often distinguished by its chronic and non-remissive nature. Early ASD diagnosis is essential because it allows for the prompt start of specialized education to successfully lessen negative impacts. However, since comorbidities are present in 10 to 25 percent of ASD patients, a correct diagnosis can be difficult and may call for further diagnostic techniques like neuroimaging. Deep learning (DL) has grown in favor of a tool to help with ASD early identification in recent years. The majority of current DL methods for ASD detection rely on Functional Magnetic Resonance Imaging (fMRI) which utilizes a small dataset and produces decent accuracy but poor generalizability. This study suggests a novel strategy to overcome this drawback and improve the efficacy of automated algorithms for autism diagnosis. In this research, a deep learning classifier based on a 1D Convolutional Neural Network (CNN) is proposed that utilizes a global filtering approach from the CCS pipeline with a widely used brain atlas, namely Craddock 200 (CC200). The study achieves better performance compared to similar existing works, with an F1-score of 88%. The observed enhancement in accuracy represents notable progress in the detection of ASD, presenting potential advantages for early intervention and assistance for individuals impacted by this condition.