Precise estimation of tissue microstructure with hybrid graph transformer.

The accurate estimation of tissue microstructure requires a sufficient amount of Diffusion MRI (DMRI) data, however, the clinical acquisition of this is challenging. Deep learning therefore improves the inference of tissue microstructure by highly undersampled DMRI. However, existing methods typically suffer from the lack of consideration of joint information in the spatial domain (x-space) and the diffusion wavevector domain (q-space). Here, we propose a hybrid graph transformer (HGT) for combined q-space learning and x-space guidance for precise estimation of tissue microstructure. The HGT consists of a q-space learning module, which explicitly considers the geometrical data structure in q-space based on a graph convolutional network, and an x-space guidance module, which learns long-range spatial dependencies based on residual dense transformer blocks. The x-space guidance module provides anatomical context to regularize the estimation of microstructure from undersampled q-space data. Extensive experiments on data from the human connectome project and high-quality diffusion-weighted imaging of Parkinson's disease indicate that HGT performs better than cutting-edge methods.