DEEP IMAGE PRIOR WITH STRUCTURED SPARSITY (DISCUS) FOR DYNAMIC MRI RECONSTRUCTION.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• High-quality training data is often unavailable for dynamic MRI.
• Reconstructing dynamic images requires capturing temporal variations effectively.

Purpose of the Study:

• To propose a self-supervised deep learning method, DISCUS, for dynamic MRI reconstruction.
• To address the challenge of limited training data in dynamic MRI.
• To discover the low-dimensional manifold of temporal variations without prior dimensionality specification.

Main Methods:

• Developed DISCUS (Deep Image Prior with Structured Sparsity), a self-supervised deep learning approach.
• DISCUS jointly optimizes network parameters and input code vectors.
• Incorporates group sparsity on frame-specific code vectors to uncover temporal dynamics.

Main Results:

• Validated DISCUS using numerical studies with dynamic phantoms and patient data.
• DISCUS successfully discovered manifold dimensionality in simulated data.
• Outperformed traditional Compressed Sensing (CS) and Deep Image Prior (DIP) in LGE phantom and patient data reconstruction.

Conclusions:

• DISCUS offers a robust method for dynamic MRI reconstruction, especially with limited training data.
• The structured sparsity approach in DISCUS enhances the discovery of true manifold dimensionality.
• DISCUS provides significant performance gains over existing methods like CS and DIP.