A Theoretical Framework for Self-Supervised MR Image Reconstruction Using Sub-Sampling via Variable Density...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Neural networks are increasingly used for reconstructing sub-sampled Magnetic Resonance Imaging (MRI) data.
• Most current methods require fully-sampled datasets for supervised training, which are often impractical to acquire.
• Self-supervised methods, utilizing only sub-sampled data, are highly desirable for broader MRI applications.

Purpose of the Study:

• To extend the Noisier2Noise framework for self-supervised reconstruction of variable density sub-sampled MRI data.
• To provide theoretical justification for the Self-Supervised Learning via Data Undersampling (SSDU) method.
• To propose and validate modifications to the SSDU method for improved performance.

Main Methods:

• Extension of the Noisier2Noise framework to address self-supervised MRI reconstruction.
• Analytical explanation of the Self-Supervised Learning via Data Undersampling (SSDU) method's performance.
• Proposal of two modifications: partitioning sampling sets and implementing loss weighting.

Main Results:

• The Noisier2Noise framework provides an analytical explanation for SSDU's effectiveness.
• Proposed modifications significantly improve SSDU's image restoration quality on the fastMRI dataset.
• Enhanced SSDU demonstrates increased robustness to partitioning parameters.

Conclusions:

• Self-supervised learning offers a viable alternative to supervised methods for MRI reconstruction when fully-sampled data is unavailable.
• The modified SSDU method, grounded in theoretical insights, advances the field of accelerated MRI.
• This research paves the way for more efficient and accessible MRI data acquisition and reconstruction.