Instance-Wise MRI Reconstruction Based on Self-Supervised Implicit Neural Representation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerated Magnetic Resonance Imaging (MRI) balances scan time with data sufficiency.
• Supervised deep learning for MRI reconstruction requires extensive fully-sampled training data, which is often impractical.
• Existing self-supervised methods may not achieve optimal reconstruction quality.

Purpose of the Study:

• To develop a novel, fully self-supervised deep learning method for MRI reconstruction.
• To enable high-quality MRI reconstruction from a single under-sampled instance.
• To reduce reliance on large, fully-sampled datasets for training.

Main Methods:

• Utilized implicit neural representation for MRI reconstruction.
• Introduced novel supervisory signals in both image and frequency domains to guide self-supervised learning.
• Trained the model using only a single under-sampled MRI instance.

Main Results:

• The proposed self-supervised method achieved superior performance compared to existing self-supervised approaches.
• The method also outperformed a conventional supervised deep learning method.
• Demonstrated strong reliability and flexibility in enhancing under-sampled MRI quality.

Conclusions:

• The novel self-supervised method significantly improves under-sampled MRI image quality.
• This approach eliminates the need for ground-truth fully-sampled images during training.
• The method offers a practical solution for accelerated MRI acquisition and reconstruction.