ASDC-Net: Optimized Convolutional Neural Network-Based Automatic Autism Spectrum Disorder Classification Using rs-fMRI Data

Abstract

Progressive neurological disorders such as Autism Spectrum Disorder (ASD) affects one in every 160 children across the country. The latest diagnostic criteria for ASD are concentrated primarily on assessing behavior. The ASD diagnostic process, on the other hand, is often time-consuming and costly, putting a significant financial burden on patient families. This paper presents an automated clinical diagnosis system for classifying ASD and typical controls (TC) using a generalized end-to-end Convolutional neural network (CNN), a deep learning (DL)-based model referred to as ASDC-Net. The proposed model provides accelerated training through Batch Normalization (BN), reducing the internal covariate shift. The resulting classification system uses functional connectivity patterns or connectomes of CC400 brain parcellation atlas of resting-state functional MRI data (rs-fMRI). The performance of the proposed model has been assessed using 1035 subjects from the Autism Brain Imaging Data Exchange (ABIDE), including 530 TC and 505 ASD. The model achieves an accuracy of (Acc) of 76.72% as compared to state-of-the-art (SOTA) methods in classifying ASD outperforming several peer machine learning (ML) and DL models. Furthermore, the Diagnostic Odd Ratio (DOR) and Matthews Correlation Coefficient (MCC) have been introduced to evaluate the effectiveness of the classifier. The evaluation metrics indicate that the suggested method is efficient and raises the prospect of helping the clinician team with the early diagnosis of ASD.