Physics-Guided Self-Supervised Implicit Neural Representation for Accelerated $\text{T}_{1\rho }$ Mapping.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Quantitative T1rho mapping is valuable for clinical and research applications but is limited by long acquisition times.
• Current deep learning methods for accelerated MRI parameter mapping often require impractical fully-sampled training datasets.
• Addressing the need for efficient and practical accelerated quantitative MRI techniques is crucial for clinical adoption.

Purpose of the Study:

• To develop a novel scan-specific, self-supervised deep learning method for accelerated quantitative T1rho mapping.
• To reconstruct T1rho-weighted images and generate T1rho maps from highly undersampled k-space data.
• To overcome the limitations of existing methods by eliminating the need for fully-sampled training data.

Main Methods:

• A self-supervised method leveraging implicit neural representation is proposed, using only spatiotemporal coordinates as input.
• The method learns an implicit neural representation guided by the physical model of T1rho mapping.
• Two explicit priors are incorporated: signal relaxation and k-t space data self-consistency.

Main Results:

• The method was validated using both retrospective and prospective undersampled k-space data.
• Achieved a high acceleration factor of up to 14, significantly reducing scan times.
• Outperformed state-of-the-art methods in artifact suppression and achieving lower reconstruction error.

Conclusions:

• The proposed scan-specific self-supervised method enables highly accelerated quantitative T1rho mapping.
• This approach overcomes the impracticality of requiring fully-sampled training data in clinical settings.
• The technique shows significant potential for improving the efficiency and applicability of T1rho mapping in MRI.