TS-SR3: Time-Strided Denoising Diffusion Probabilistic Model for MR Super-Resolution
Abstract: Iterative refinement based image super-resolution with conditional denoising diffusion probabilistic models (DDPM), such as SR3 [21], has shown promise in the super-resolution of magnetic resonance images (MRIs). However, these methods are dependent on the inference stage of the DDPMs, which can be slow and also require hundreds of iterations to reach the desired denoising level. We address this issue by proposing a time-strided SR3 (TS-SR3) for MRI super-resolution. Traditional DDPM approaches add noise to the high-resolution (HR) image according to small steps of the variance schedule, and the accumulation of the noise results in an image resembling pure Gaussian noise. In contrast, we take larger strides in time across the same variance schedule, leading to less accumulation of noise and diffusing to a different overall noise level. At inference time, this permits us to start denoising not from full noise but with some signal still present. We propose three ways in which to generate the initial estimate where the signal is still present and evaluate the benefits of each. Our experiments show that our TS-SR3 approach is superior to a recently published super-resolution method and that our two alternative initialization approaches further improve results.
Loading