Autism Screening Based on Deep Learning Analysis of Eye Movement Trajectories

Traditional clinical diagnosis of autism spectrum disorder (ASD) relies heavily on subjective physician assessments, resulting in longer diagnostic cycles and low efficiency. Eye‐tracking technology offers a noninvasive approach to objectively record visual attention patterns, enabling the identification of autism‐specific visual behavioural features that serve as diagnostic indicators, showing unique advantages for autism diagnosis. This study proposes ET‐ASDNet, a deep learning framework based on ResNet18 with three specialised modules for analysing eye movement trajectories to assist in ASD screening. The framework incorporates dynamic trajectory perception, multi‐dimensional cross‐attention fusion and adaptive spatial feature integration to capture ASD‐specific visual patterns. Evaluation was conducted using stratified three‐fold cross‐validation on two public eye‐tracking datasets: the Eye‐tracking Scanpath Dataset, where ET‐ASDNet achieved 85.19% accuracy and 86.96% F1‐score, and the Saliency4ASD dataset, where it achieved 72.67% accuracy and 76.84% F1‐score. Results demonstrate that ET‐ASDNet can provide an objective tool for auxiliary ASD screening.