S³CNet: self-supervised Siamese cooperative network for accelerating magnetic resonance imaging reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning (DL) accelerates MRI acquisition but requires extensive fully sampled data.
• Clinical MRI data limitations hinder the practical application of DL reconstruction methods.
• Current DL methods for MRI reconstruction face challenges due to the scarcity of fully sampled labeled data.

Purpose of the Study:

• Propose a self-supervised Siamese cooperative network (S3CNet) for high-quality MRI reconstruction.
• Reduce reliance on fully sampled data for DL-based MRI reconstruction.
• Maintain excellent reconstruction performance using undersampled k-space data.

Main Methods:

• Developed a two-stage self-supervised reconstruction framework (S3CNet) compatible with various sampling patterns.
• Employed measurement consistency and cross-consistency losses for noise suppression in the first stage.
• Utilized reconstruction consistency loss with pseudo-labels in the second stage for enhanced generalization.

Main Results:

• S3CNet demonstrated superior performance compared to state-of-the-art self-supervised methods on the FastMRI dataset.
• Reconstruction quality achieved by S3CNet was comparable to fully supervised methods.
• The method excelled across diverse undersampling rates and sampling patterns.

Conclusions:

• S3CNet effectively eliminates the need for fully sampled training data in DL-based MRI reconstruction.
• The network achieves high reconstruction quality and strong generalization capabilities.
• Presents a practical solution for accelerating clinical MRI acquisition.