Learning Disentangled Representation for Multidimensional MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Multidimensional magnetic resonance imaging (MRI) generates complex datasets.
• Current reconstruction methods face challenges in disentangling various image features.
• Accelerated MRI acquisition is crucial for clinical efficiency.

Purpose of the Study:

• To propose a novel representation and reconstruction framework for multidimensional MRI.
• To enable disentanglement of image features like contrast and geometry.
• To improve the efficiency and accuracy of MRI reconstruction, particularly for accelerated protocols.

Main Methods:

• An autoencoder was trained to learn disentangled latent spaces for feature representation.
• A latent diffusion model was employed to capture distributions of disentangled features.
• A new formulation integrated learned representations with complementary constraints for reconstruction from sparse data.

Main Results:

• Successfully disentangled contrast and geometry features in multicontrast MRI.
• Demonstrated effectiveness in accelerated T1 and T2 mapping.
• The proposed method enables robust reconstruction from sparse data.

Conclusions:

• The developed method offers a powerful approach for multidimensional MRI reconstruction.
• Feature disentanglement enhances the representation of complex MRI data.
• This technique holds promise for accelerating MRI acquisition and improving diagnostic capabilities.