TinyDINO VHaar based Bi-directional Chameleon optimized LightASDNet for autism spectrum disorder detection

Area of Science:

• Neuroscience
• Computer Science
• Developmental Pediatrics

Background:

• Early detection of Autism Spectrum Disorder (ASD) in toddlers is crucial but challenging due to neurodevelopmental deficits and variability.
• Current diagnostic methods lack specific biomarkers, hindering timely intervention.

Purpose of the Study:

• To propose an advanced AI framework, TiDI-ASDNet, for accurate and early detection of ASD in toddlers.
• To integrate visual and linguistic data processing for a comprehensive diagnostic approach.

Main Methods:

• Utilized Optimized Hierarchical Guided Image Filter (OHGF) for image preprocessing and DINOv2-NetVLAD for visual pattern extraction.
• Employed One-Hot SMOTE (OHS) for text preprocessing and a Hybrid Tiny Encoder based HaarNet (TE-HNet) for linguistic feature extraction.
• Implemented Bi-directional Encoder based Cross Factorization (BiE-xF) to reduce ambiguity and Modified Chameleon optimized LightASDNet (MC-LAN) for final classification.

Main Results:

• The TiDI-ASDNet framework demonstrated high performance in classifying ASD and non-ASD cases.
• Achieved an Area Under the Curve (AUC) of 99.2% and an accuracy of 98.2% in simulations.
• The model effectively mitigated challenges like motor stimming and linguistic complexities.

Conclusions:

• The proposed TiDI-ASDNet framework offers a robust and highly accurate solution for early ASD detection in toddlers.
• This AI-driven approach shows significant potential in improving diagnostic outcomes and enabling timely interventions.