SelfCoLearn: Self-Supervised Collaborative Learning for Accelerating Dynamic MR Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning accelerates dynamic magnetic resonance (MR) imaging.
• Current methods struggle with fine details due to limited training data.
• Need for accurate reconstruction from undersampled k-space data.

Purpose of the Study:

• Propose a self-supervised collaborative learning framework (SelfCoLearn).
• Enable accurate dynamic MR image reconstruction from undersampled k-space data.
• Improve recovery of fine details and structures in dynamic MR imaging.

Main Methods:

• Developed a self-supervised collaborative learning framework (SelfCoLearn).
• Incorporated dual-network collaborative learning, re-underampling data augmentation, and a co-training loss.
• Integrated the framework into model-based iterative un-rolled networks.

Main Results:

• SelfCoLearn demonstrated strong capabilities in direct reconstruction from undersampled k-space data.
• The method effectively captures essential and inherent image representations.
• Achieved high-quality and fast dynamic MR imaging.

Conclusions:

• SelfCoLearn offers a robust solution for accurate dynamic MR image reconstruction.
• The framework enhances image quality and reconstruction speed.
• Addresses limitations of existing methods by eliminating the need for fully sampled reference data.