Autism Spectrum Disorder Identification Using Dual-Branch Fusion Model with Privacy-Preserving

Abstract

Autism Spectrum Disorder (ASD) are neurodevelopmental disorders that severely impact daily life and social interactions. According to research, early diagnosis and intervention of autism is crucial to improve the overall quality of life of patients. Although existing machine learning and deep learning methods have been applied to the identification and detection of autism, healthcare organizations often refuse to share or disclose medical data with the improvement of laws and regulations. Therefore, we propose a privacy-preserving deep learning method based on the local client using a dual-stream model to further improve the ASD recognition performance by capturing the features of functional MRI in both temporal and spatial structures, and further ensure that each client improves the performance of the local recognition task through federated learning by optimizing the two steps of the local client update and the client aggregation during federated learning. The experimental results show that our model achieves the best AUC of 0.952, which ensures the overall performance of the classification model, and the recognition accuracy is significantly improved by using federated learning compared to the results when clients are trained independently.