A Novel Hybrid Vision Transformer�XGBoost Framework for Automated Autism Spectrum Disorder Detection Using Image-Based Data

Abstract

Our study focusses on a novel deep learning framework that integratesVision Transformer-based Autism Spectrum Disorder detection usingViTASD and DeiT along with XGBoost to enhance classificationperformance using image-based data. ASD is a multifaceted disease andtheearly diagnosis has been a serious challenge in terms of disease detection. It is a neurodevelopmental disorder in which precise diagnosis is oftenhampered due to its varied characteristics and dependence on subjectiveevaluations. Conventional diagnostic methods, such as behavioral checklists and clinical assessments, are labor-intensive, necessitate expert participation, and may lack applicability across different demographicgroups. The suggested hybrid model has been pre-trained on a ResNet50architecture and utilizes the feature extraction strengths of gradient boostedtress (XGBoost) combined with the attention-focused accuracyof transformer architecture of ViTASD and DeiT to detect distinctive patternsthat may be linked to ASD. This model has been trained and validatedondifferent datasets, and its efficiency was compared to existing machinelearning models. Our findings have exhibited a notable enhancement inaccuracy, specificity, and overall robustness highlighting the model’spotential as a reliable, scalable, and serviceable screening instrument. Thisstudy also addresses some shortcomings of previous research by providinga new approach for future studies regarding ASD detection, withencouraging prospects for both clinical and remote diagnostic applications. Our study has shown significant results with an accuracy of 97%andgoodoverall performance metrics which suggests a successful predictive analysisof ASD detection.