High-Frequency Conditional Diffusion Probabilistic Model for Under-sampled MRI Reconstruction

The magnetic resonance imaging (MRI) is extensively used in medical clinical diagnosis due to its excellent performance in medical imaging. However, the disadvantage of slow imaging speed has hindered the further application of MRI in medical clinical diagnosis. Diffusion probabilistic models (DPM) have achieved impressive results in image restoration due to the extraordinary ability to generate highly realistic images. However, DPM needs to be set up with a long number of diffusion steps to obtain excellent results, leading to the training and sampling process to consume a lot of time. This paper presents a high-frequency conditional diffusion Probabilistic Model (HFC-DDPM), which defines the diffusion and inference process in wavelet domains rather than image domain. On this basis, the Wave-DiffNet, a network for solving MRI reconstruction tasks is presented, which recovers the high-frequency and low-frequency components of MR image by diffusion model and encoder-decoder structure, respectively. Notably, to prevent conflicts between frequencies caused by optimizing objectives at different stages, the Frequency Cross-attention Alignment module is designed to align and normalize the frequencies at each iteration of the denoising process. Our model has been evaluated on the MICCAI Grand Challenge dataset and has significantly improved the quality and speed of MRI image reconstruction.