Phase-constrained zero-shot self-supervised learning for BLADE liver MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Liver MRI is crucial for disease diagnosis and monitoring.
• Image quality is often compromised by motion and noise, especially in high-resolution and diffusion-weighted imaging.
• Propeller-based sequences like BLADE enhance motion robustness but require advanced reconstruction for accelerated imaging.

Purpose of the Study:

• To develop and evaluate a novel self-supervised learning framework for accelerated BLADE liver MRI reconstruction.
• To improve image quality by reducing artifacts and noise without external training data.

Main Methods:

• A phase-constrained zero-shot self-supervised learning (PC ZS-SSL) framework was proposed, embedding BLADE operators in a deep network with phase estimation.
• The method utilizes partitioned k-space for self-supervised learning, eliminating the need for external training datasets.
• Evaluated using phantom experiments and in vivo T2-weighted and diffusion-weighted imaging, compared against LLR and vendor reconstructions.

Main Results:

• PC ZS-SSL demonstrated superior preservation of fine details and sharpness in phantom studies compared to LLR.
• In vivo, it effectively reduced noise and ringing artifacts in T2W and DWI scans, maintaining anatomical fidelity.
• Achieved comparable or superior image quality to LLR and vendor methods, particularly in high-noise DWI and under severe undersampling conditions.

Conclusions:

• PC ZS-SSL enables high-quality, artifact-suppressed BLADE liver MRI reconstruction without external training data.
• The framework shows robust performance, especially in challenging diffusion-weighted imaging scenarios.
• Highlights potential for clinical translation in improving liver MRI diagnostics.