SELFIE: Self-Supervised Learning for Fast Dynamic Golden-Angle Radial MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Accelerated golden-angle radial MRI is crucial for dynamic imaging.
• Compressed sensing (CS) and supervised deep learning (DL) have limitations in reconstruction speed and reference data requirements.

Purpose of the Study:

• To introduce SELFIE (Self-supervised Learning for Fast dynamic golden-Angle radial MRI), a novel self-supervised reconstruction technique.
• To enable fast and high-quality dynamic MRI reconstruction without requiring fully-sampled or CS training references.

Main Methods:

• SELFIE utilizes slice-by-slice image-domain processing, avoiding k-space computations.
• Self-supervision is achieved through variable temporal resolution and temporal sparsity inherent in golden-angle radial sampling.
• The method was validated on dynamic contrast-enhanced (DCE)-MRI and motion-resolved abdominal MRI datasets.

Main Results:

• SELFIE demonstrated image quality and dynamic fidelity comparable to CS and supervised DL.
• Reconstruction time was significantly reduced, with 3D time series reconstructed in seconds.
• Reader studies showed SELFIE performed comparably to supervised DL and better than CS.

Conclusions:

• SELFIE offers a fast, reference-free reconstruction framework for dynamic golden-angle radial MRI.
• It presents a viable alternative to existing CS and supervised DL methods.
• The technique achieves competitive performance and substantially improved reconstruction speed.