VMamba UNet with Latent Diffusion Prior Guidance for MRI Reconstruction

Motivation: Latent space diffusion models1 can efficiently generate high-quality images. However, the inherent nonlinearity and nonconvexity of variational autoencoder decoders present challenges for their application to MRI reconstruction. Goal(s): Effectively integrating both latent and pixel space information to guide MRI reconstruction indirectly. Approach: We propose a VMamba2 UNet that employs a VMamba-based encoder to align latent and pixel space prior guidance, along with a convolutional decoder. Additionally, we introduce Latent Guided Attention (LGA) to effectively integrate information from both latent and pixel spaces. Results: Experiments demonstrate that our method achieves superior performance compared to the baselines. Impact: This work investigates the role of generative priors of latent diffusion models on MRI reconstruction tasks and facilitates the frontier studies of deep-learning-based rapid reconstruction methods.