Repeatability-encouraging self-supervised learning reconstruction for quantitative MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Quantitative MRI

Background:

• Quantitative MRI (qMRI) clinical value depends on measurement repeatability.
• Deep learning (DL) accelerates qMRI reconstruction but often neglects repeatability.
• Existing DL methods may compromise quantitative accuracy for speed.

Purpose of the Study:

• To develop a repeatability-encouraging self-supervised learning (SSL) reconstruction method for quantitative MRI.
• To improve the clinical utility of accelerated quantitative MRI techniques.
• To address the trade-off between reconstruction speed and quantitative accuracy in DL-based qMRI.

Main Methods:

• A novel SSL reconstruction network was designed, minimizing cross-data-consistency between k-t-space subsets.
• The network's repeatability was promoted by enabling reconstructions to predict complementary k-t-space data.
• Cardiac MR Multitasking T1 mapping data were used for evaluation, comparing SSL with supervised learning (Sup60, Sup30/30).

Main Results:

• The SSL method achieved image and T1 map quality comparable to supervised methods trained on full data (Sup60).
• SSL demonstrated superior T1 repeatability (6.3% CV) compared to Sup60 (12.0%) and Sup30/30 (6.9%) on split data.
• SSL reconstructions were sharp and repeatable, outperforming supervised methods in quantitative agreement and image quality.

Conclusions:

• The proposed SSL method offers superior repeatability and sharpness in quantitative MRI reconstruction without labeled data.
• SSL provides a significant advancement over supervised learning methods for accelerated qMRI.
• This approach enhances the clinical feasibility of quantitative MRI by improving reliability and efficiency.