Microstructure Estimation Using Synergistic Dual-path Hybrid Network

White matter microstructure plays a pivotal role in the diagnosis and study of brain disorders. Deep learning-based estimation of white matter microstructural indices from diffusion MRI (dMRI) data has gained increasing research attention in recent years. To conduct effective learning in the heterogeneous space (i.e., x-space and q-space) of dMRI data, hybrid neural networks were proposed and have shown great potential. To this end, we propose a new hybrid neural network, called synergistic dual-path hybrid convolutional neural network (SDH-Net), for more effective microstructure estimation. In our SDH-Net, we propose a dual-path architecture that takes bidirectional asymmetric learning in x-space and q-space to enhance feature representation in heterogeneous domains. Firstly, 3D patches are extracted from each diffusion gradient as vertices, and then a graph is constructed based on the correlation between diffusion angles. In the x-q learning branch, an effective representation is embedded in x-space to enhance learning in q-space, while vice versa in the q-x learning branch. Extensive experiments on data from the human connectome project demonstrate that our SDH-Net outperforms the existing state-of-the-art models.