Consistent Direct Diffusion Bridge with Injection for MRI Reconstruction

Pyatkovskiy Sergey, Jaejun Yoo

Keywords: Magnetic Resonance Imaging, Diffusion Bridge, Reconstruction

Abstract: In MRI, lengthy acquisition times often cause motion artifacts. We propose a new method, Consistent Direct Diffusion Bridge with Injection (CCDBI), which leverages diffusionbased image priors to reconstruct MRI from undersampled k-space data. Unlike traditional diffusion methods starting with Gaussian noise, CCDBI begins sampling from actual measurements, improving accuracy by aligning with the target image. By combining information from noisy image domain and k-space adaptively, CCDBI ensures consistency in the Direct Diffusion Bridge (DDB), enhancing reconstruction quality. Experimental results on IXI and OASIS-2 datasets demonstrate CCDBI’s superiority over existing algorithms.

Submission Number: 169