DALNN: Dual Attention Learning Neural Network for Diagnosis of Autism Spectrum Disorders and Exploration of Lesion Brain Regions

Autism spectrum disorders (ASDs) typically occur in early childhood and affect brain development. Identifying the diseased regions of interest (ROIs) is crucial for improving early diagnosis. However, most current deep learning algorithms for ASD diagnosis struggle to locate lesion ROIs and fully explore relationships between them, which limits their clinical applicability. To resolve these, a dual attention learning fusion algorithm (DALFA) is proposed to enhance feature extraction and fusion from multimodal data. Specifically, the algorithm first constructs features based on both the functional and structural data of the subjects and applies appropriate methods for feature extraction. After concatenating the extracted features, the dual attention learning is employed to fuse these features across two dimensions, optimizing feature representation. Furthermore, this study introduces a dual attention learning neural network (DALNN) that implements and refines the details of DALFA. We conduct sufficient validation experiments using data from the ABIDE dataset. The classification accuracy of DALNN reaches 83.35%, outperforming existing methods. In addition, DALNN can extract ASD-related ROIs, which is beneficial for guiding early clinical intervention.