Time-Conditioned Zero-Shot Self-Supervised Reconstruction for Accelerated 3D Ultra-Low-Field MRI.

Area of Science:

• Magnetic Resonance Imaging
• Medical Imaging Technology
• Artificial Intelligence in Medicine

Background:

• Ultra-low-field (ULF) MRI offers a cost-effective and portable alternative to conventional MRI.
• However, ULF MRI suffers from lower signal-to-noise ratio (SNR) and longer scan times.
• Accelerated image acquisition is crucial for ULF MRI's clinical viability.

Purpose of the Study:

• To develop a training-free image reconstruction framework for accelerating 3D single-coil ULF MRI.
• To introduce a transfer-learning variant for even faster computation.
• To evaluate the proposed methods' performance and generalizability.

Main Methods:

• A time-conditioned zero-shot self-supervised learning (ULF-ZS-SSL) framework was developed.
• This method combines physics-based data consistency with a 3D residual network and time-step embeddings.
• A transfer-learning variant (ULF-ZS-SSL-TL) was created by pretraining on a small ULF brain dataset.

Main Results:

• ULF-ZS-SSL achieved high-quality reconstructions across various contrasts, outperforming traditional methods.
• Time-step conditioning enhanced convergence speed.
• ULF-ZS-SSL-TL accelerated reconstruction threefold, enabling 3D scans in ~3 minutes and showing cross-anatomy generalization.

Conclusions:

• The ULF-ZS-SSL framework allows accurate, training-free reconstruction of undersampled ULF MRI data.
• The ULF-ZS-SSL-TL approach further speeds up reconstruction using minimal training data.
• This technology supports rapid and robust applications in portable or resource-limited ULF MRI systems.