Generating missing M0 in ADNI ASL dataset with latent diffusion models

Motivation: Alzheimer’s Disease Neuroimaging Initiative (ADNI) ASL dataset acquired on Siemens scanners missed M0 images, which prevents CBF quantification for further analysis. Goal(s): Our goal is to generate the missing M0 for the ADNI ASL dataset using latent diffusion model (LDM). Approach: A separate training dataset was acquired with the ADNI ASL protocol but with manually disabled background suppression to be used as the M0. A conditional LDM was trained to use acquired control images as the condition to generate M0 images. Results: The generated M0 with the conditional LDM shows a high fidelity compared to the experimentally acquired M0. Impact: With generated M0 images, more than 500 ADNI ASL datasets can be further analyzed for CBF to investigate AD progression.