Learning temporal characteristics in multi-contrast MR images with self-supervision: An application to accelerating quantitative T2 mapping

Motivation: Accurate quantification for parameter mapping requires sufficient sampling of temporal signal evolution. Current DL-based approaches to learn parameter maps with fewer multi-contrast images often rely on fixed input parameters, limiting their flexibility Goal(s): To learn temporal characteristics of underlying tissues in multi-contrast MR images to provide a flexible DL model for accelerated quantitative T2-mapping. Approach: A vision transformer (T2-ViT) is combined with masked auto-encoder training to learn model-free T2 signal evolution given random temporal under-sampling. Results: Given the first three TE images, the model can predict T2w images at longer TE times with high structural similarities and low T2-estimation errors, making acceleration possible. Impact: An understanding of underlying temporal characteristics of tissues with vision transformers can help with intelligent design of current multi-contrast data acquisition schemes.