Motion-Guided Deep Image Prior for 3D Real-Time Cine (M-DIP-3D)

Motivation: Three-dimensional real-time (3D-RT) cardiovascular magnetic resonance imaging (CMR) requires extremely high acceleration rates, which are not feasible using existing methods. Goal(s): Develop an unsupervised reconstruction method for 3D-RT cine to capture beat-to-beat variations in the cardiac function. Approach: Our method, termed M-DIP-3D, builds upon the deep image prior (DIP) framework and incorporates manifold learning to model motion and content variation. Results: For a 3D cine MRXCAT phantom with both regular and irregular beats, M-DIP-3D effectively captures the beat dynamics. We also demonstrate the feasibility of M-DIP-3D for ferumoxytol-enhanced 3D-RT cine imaging in a healthy subject and a patient. Impact: The proposed method, M-DIP-3D, facilitates 3D real-time cine imaging from highly undersampled data. By directly learning the underlying motion and content variation manifold, M-DIP-3D produces images with minimal motion blur and real-time dynamics.