Accelerated Patient-specific Non-Cartesian Magnetic Resonance Imaging Reconstruction Using Implicit Neural...

Area of Science:

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

Background:

• Accelerated MRI is crucial for image-guided therapies.
• Compressed Sensing (CS) and deep learning methods like CNNs/Transformers have limitations in modeling continuous k-space, especially with non-Cartesian sampling.
• Implicit Neural Representations (INRs) can model continuous signals in the frequency domain, making them suitable for arbitrary k-space sampling.

Purpose of the Study:

• To develop novel k-space generative-adversarially trained INRs (k-GINR) for de novo undersampled non-Cartesian k-space reconstruction.
• To evaluate the performance of k-GINR against existing reconstruction methods.

Main Methods:

• k-GINR employs a two-stage approach: supervised training on a patient cohort followed by self-supervised patient-specific optimization.
• The StarVIBE T1-weighted liver dataset (118 scans) was used for testing.
• k-GINR was compared with INR-based methods (NeRP, k-NeRP), Deep Cascade CNN, and CS.

Main Results:

• k-GINR demonstrated superior performance over baseline methods, particularly at high accelerations (e.g., 29.3%-60.5% higher PSNR at 20x acceleration).
• Reconstruction times were competitive, with k-GINR at approximately 3 minutes, comparable to NeRP and k-NeRP, and faster than CS and Deep Cascade CNN in some aspects.
• Significant improvements in peak-signal-to-noise ratio (PSNR) were observed across various acceleration factors.

Conclusions:

• k-GINR is an innovative two-stage INR network with adversarial training for direct non-Cartesian k-space reconstruction.
• The method achieves superior image quality compared to CS and Deep Cascade CNN, even at high acceleration ratios.
• k-GINR offers a promising solution for accelerated MRI reconstruction in clinical settings.