Dual-domain self-supervised learning for accelerated non-Cartesian MRI reconstruction

Area of Science:

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

Background:

• Current deep MRI reconstruction networks require fully sampled data and are limited to Cartesian patterns, hindering clinical adoption.
• Non-Cartesian sampling offers advantages in acceleration and motion robustness but presents reconstruction challenges.
• Supervised deep learning methods necessitate extensive, fully sampled data, which is often impractical for clinical MRI acquisition.

Purpose of the Study:

• To introduce a fully self-supervised approach for accelerated non-Cartesian MRI reconstruction.
• To overcome limitations of supervised methods and enable practical adoption of accelerated MRI.
• To leverage self-supervision in both k-space and image domains for improved reconstruction accuracy.

Main Methods:

• Developed DDSS (Deep Dual Self-Supervision), a novel self-supervised deep learning framework.
• Utilized disjoint k-space partitions for k-space self-supervision, training the network to reconstruct data from subsets and itself.
• Enforced image-level self-supervision by ensuring appearance consistency between original undersampled data and reconstructed partitions.

Main Results:

• DDSS achieved high-quality reconstructions on simulated multi-coil non-Cartesian MRI data, approaching fully supervised accuracy.
• Outperformed existing baseline methods in reconstruction quality on simulated datasets.
• Successfully scaled to challenging real-world clinical MRI data from a portable low-field scanner without supervised training data.
• Demonstrated improved image quality compared to traditional reconstruction methods in a radiologist study.

Conclusions:

• DDSS offers a powerful self-supervised solution for accelerated non-Cartesian MRI reconstruction.
• The method effectively handles undersampled data and achieves high reconstruction accuracy without fully sampled training data.
• DDSS shows significant potential for clinical translation, particularly in scenarios with limited data availability or challenging acquisition conditions.