Accelerating DWI Data Acquisition: Reconstruction Using Graph-Driven Denoising Diffusion Probabilistic Model

To reduce time cost in acquiring diffusion weighted imaging (DWI) data, we propose a novel graph-driven denoising diffusion probabilistic model (gDDPM) to reconstruct DWI volumes from ones undersampled at a rate of four, by taking into account their q-space correlations. gDDPM is composed of two main modules, i.e., 1) b0-guided reconstruction module, which employs a fully sampled b0 volume to provide anatomical constraints for reconstructing low-SNR and varying SNR volumes in a diffusion model, and 2) q-space consistency module, which introduces a pretrained graph neural network encoder to maintain q-space consistency of the reconstructed volumes. Extensive experiments on an in-house dataset have demonstrated the superiority of our proposed gDDPM over three state-of-the-art methods. In addition, gDDPM serves as a compelling example, demonstrating the potential of integrating both b0 guidance and q-space consistency constraints in DWI data reconstruction tasks. The codes will be available at https://github.com/zhongruiHuangDMRI/GDDPM.