Orthogonal tensor dictionary learning for accelerated dynamic MRI.

Area of Science:

• Medical Imaging
• Signal Processing
• Machine Learning

Background:

• Dynamic MRI data is inherently multidimensional (tensor), but traditional methods vectorize/matricize it, losing spatial structure.
• Compressed Sensing (CS) theory applied to dynamic MRI reconstruction often uses 1D/2D models, which can degrade data properties.
• Tensor decomposition offers a way to exploit multidimensional data structure for improved MRI reconstruction.

Purpose of the Study:

• To develop a novel tensor dictionary learning method for dynamic MRI reconstruction.
• To improve reconstruction accuracy and computational efficiency compared to existing methods.
• To preserve the inherent spatial structure of dynamic MRI data during reconstruction.

Main Methods:

• Introduced a novel tensor dictionary learning method with an orthonormal constraint on the tensor dictionary's elementary matrix.
• Developed an algorithm that alternates sparse coding, tensor dictionary learning, and reconstruction updating.
• Solved each subproblem efficiently using a closed-form solution.

Main Results:

• The proposed tensor dictionary learning method demonstrated significant improvements in reconstruction accuracy.
• The new scheme achieved better computational efficiency compared to the traditional 1D/2D model with overcomplete dictionary learning.
• Numerical experiments on phantom and synthetic data validated the effectiveness of the proposed method.

Conclusions:

• The novel tensor dictionary learning approach effectively reconstructs dynamic MRI data while preserving its multidimensional structure.
• This method offers a superior alternative to traditional 1D/2D models for dynamic MRI reconstruction, enhancing both accuracy and speed.
• The orthonormal constraint and efficient closed-form solutions contribute to the method's practical applicability.