Keywords: MRA, diffusion probabilistic model, high frequency signal, synthesis, vasculature
Abstract: Magnetic resonance angiography (MRA) allows for the non-invasive visualization of vascu- lature in the human body and has been widely used in the hospitals to identify aneurysms and the location of a stroke. Generating MRA using the commonly available T1-weighted (T1w) MRI modality would broaden the possibilities for studying vasculature because T1w is commonly acquired in most neuroimaging datasets, while MRA is not. In this work, we propose a method using the statistical generative model called denoising diffusion prob- abilistic model (DDPM) to tackle the MRA synthesis task. Our experiment shows that by diffusing the high frequency signal, which explains the major signal difference between MRA and T1w, DDPM could successfully synthesize MRA with good quality. The pro- posed method also conditioned score-matching estimation with the high frequency signal of the T1w modality, which enables the accurate one-to-one synthesis between MRA and T1w.