Exploring latent space representations of T1/T2 relaxation, cardiac motion, and respiratory motion for multidimensional quantitative CMR

Motivation: Respiratory and cardiac motion identification is challenging with changing contrast weightings for self-gated multidimensional techniques like MR multitasking. Goal(s): To guide VAE latent vector constraints design for representing relaxation and motion. Approach: We evaluated VAE representational fidelity for 16 combinations of constraints on T1/T2 relaxation, cardiac, and respiratory latent dimensions. Results: The results demonstrate that nonlinear T1/T2 relaxation representations and cardiac phase representations improve VAE performance. Impact: Latent space design is important for VAE learning in multidimensional cardiac imaging, suggesting avenues for better self-gated cardiac and respiratory binning.