ASD-Mamba: Mamba-based prediction of atypical visual saliency in autism spectrum disorder

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by atypical visual attention patterns, which are frequently manifested as gaze anomalies. Although significant advancements have been made in visual attention modeling in terms of both accuracy and efficiency, most existing methods are primarily designed for predicting typical attention patterns and exhibit limited generalization capability when applied to individuals with ASD. This constraint impedes further progress in the prediction of atypical visual attention. To address this issue, we propose ASD-Mamba, a novel framework that integrates the efficient long-range dependency modeling capability of Mamba with a U-Net architecture for predicting atypical visual saliency in ASD. By incorporating an Efficient Adaptive Visual State Space (EA-VSS) block, ASD-Mamba improves feature extraction. The Frequency Domain Attention (FDA) module reduces spatial and temporal complexity while retaining essential feature information. Additionally, the Conditional Attention Fusion (CAF) module effectively preserves more information during feature fusion, thereby ensuring richer contextual representations. Experimental results on an ASD eye-tracking dataset demonstrate the competitive performance of our model in atypical visual saliency prediction for children with autism. Our code and pretrained models are available athttps://github.com/ASDproject1/ASD-Mamba.