Multidimensional MR Image Reconstruction Using A Disentangled Representation

Motivation: Learning-based high-dimensional image reconstruction may benefit from pre-training strategies and representations better exploiting multidimensional correlations. Goal(s): To develop a disentangled representation that can effectively model and constrain different types of features in multidimensional images for reconstruction. Approach: The representation was learned via image transfer autoencoder training. A new formulation was proposed to incorporate the pre-learned representation for constrained reconstruction. Results: Our model can disentangle contrast and geometry features in multi-contrast MR images and the proposed algorithm improved quantitative MRI reconstruction. Impact: The proposed method may provide a new perspective for learning-based, high-dimensional MRI reconstruction, for which small or even no data are available problem-specific supervised training.