Machine Learning-Powered Brain Imaging for Early Detection of Autism Spectrum Disorder

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition whose accurate and early detection remains an enduring challenge in clinical neuroscience. Resting-state functional magnetic resonance imaging (rs-fMRI) provides a non-invasive means of probing functional brain connectivity, yet multi- site variability, class imbalance, and inconsistent methodological practices

continue to hinder progress. In this thesis, we conduct a systematic benchmarking study on the merged ABIDE-I and ABIDE-II datasets, comprising more than 2,200 subjects from over 30 international sites. Fourteen machine learning models— encompassing linear classifiers, probabilistic approaches, ensemble methods, boosting algorithms, and neural networks—were implemented under a unified preprocessing and evaluation pipeline. Performance was rigorously assessed using accuracy, precision, recall, and F1-macro to ensure robustness against dataset imbalance. Results demonstrate that ensemble and boosting approaches, particularly Random Forest, LightGBM, Gradient Boosting, and CatBoost, consistently outperform classical baselines and rival deep learning methods, achieving accuracies approaching 98% and F1-macro scores exceeding 0.99. This study establishes a reproducible baseline for ASD classification, highlights interpretable and deployment-ready models for potential clinical decision-support, and lays the groundwork for advancing towards deep, graph-based, and multimodal neuroimaging frameworks that address the complexity of ASD.