Interpolating Dynamic MRI Images with a Latent Brownian Bridge Diffusion Model

Motivation: The acquisition time of dynamic magnetic resonance is long, and the existing interpolation results are not fine enough. Goal(s): Develop a new interpolation model that can accurately insert the intermediate frame between the input frames. Approach: We propose a latent Brownian bridge diffusion model. The model introduces Brownian bridge diffusion to achieve a much smaller cumulative change in the generated latent representation. Results: We tested on the private data set and got a PSNR of 40.12, and on the public data set ACDC we got a PSNR of 34.37. The interpolation effects are both better than the compared models. Impact: The potential Brownian Bridge diffusion model can predict the intermediate frame with high efficiency and high quality thanks to the certainty of diffusion. At the same time, our model is the first application of LDM architecture in medical image interpolation.