IMJENSE: Scan-Specific Implicit Representation for Joint Coil Sensitivity and Image Estimation in Parallel MRI.

Area of Science:

• Medical Imaging
• Computational Physics
• Artificial Intelligence

Background:

• Parallel imaging accelerates Magnetic Resonance Imaging (MRI) acquisition by using undersampled k-space data.
• Reconstructing high-quality MRI images from highly undersampled data remains a significant challenge.
• Implicit neural representation offers a novel approach to leverage data physics for improved reconstruction.

Purpose of the Study:

• Introduce IMJENSE, a scan-specific implicit neural representation method for advanced parallel MRI reconstruction.
• Model MRI images and coil sensitivities as continuous functions using neural networks and polynomials.
• Enable direct learning from undersampled k-space data without requiring fully sampled ground truth.

Main Methods:

• Developed IMJENSE, an implicit neural representation-based technique for parallel MRI.
• Parameterized MRI images and coil sensitivities as continuous functions of spatial coordinates.
• Simultaneously learned neural network weights and polynomial coefficients directly from sparse k-space measurements.

Main Results:

• IMJENSE demonstrated superior performance compared to conventional and deep learning-based reconstruction algorithms.
• Achieved robust image reconstruction at 5x and 6x accelerations with limited calibration lines (4 or 8).
• Showcased stability with extremely limited calibration data, outperforming supervised methods.

Conclusions:

• IMJENSE provides high-quality, scan-specific parallel MRI reconstruction.
• The method holds significant potential for further accelerating MRI data acquisition.
• Implicit neural representation offers a powerful paradigm for addressing inverse problems in MRI.