Diff2-SRNet: Super-Resolved Q-space Learning in Diffusion MRI using Non-Gaussian Residual Conditional Diffusion Model

Motivation: The scanning time of high-angular resolution diffusion MRI increases linearly with the number of diffusion gradients, which limits its widespread use in clinical settings. Goal(s): We aimed to facilitate high-fidelity and detail-preserving super-resolution for dMRI in q-space. Approach: We propose a physical knowledge-guided residual DDPM-based method, Diff2-SRNet. This model divides the dMRI signal into Gaussian and non-Gaussian components, then utilizes the diffusion tensor model to represent the former part and leverages the strong generative capabilities of DDPM for the more complex non-Gaussian component. Results: Experimental results demonstrate that the proposed Diff2-SRNet reconstructs HAR DWIs with higher fidelity and preserves better details. Impact: The proposed method exhibits interpretability and reliability and shows a high potential to become a practical tool in a wide range of clinical and neuroscientific applications.