Self-Supervised SUper-Resolution ASL Enhancement based on 3D Latent Diffusion Models (SURED-L)

Motivation: Arterial Spin Labeling (ASL) imaging suffers from low SNR, low resolution, and long acquisition times, hindering its clinical applications. Goal(s): To propose a self-supervised ASL super-resolution framework that utilizes a 3D latent and image-space diffusion model. Approach: A 3D latent space conditional diffusion model was trained using multimodal images, including T1w and ASL. The ASL super-resolution model leverages latent space information from T1w. The method was tested on ASLs acquired at low and high resolutions. Results: The proposed model provides super-resolution ASL with enhanced details, improved SNR, high visual scores. It was more efficient than the previous ASL diffusion model. Impact: The proposed method achieved ASL super-resolution by combining latent and image-space models, which can enhance the resolution of 4mm ASL to 2.5mm, equivalent to reducing the scan time by 13 mins.