Noise2Recon: Enabling SNR-robust MRI reconstruction with semi-supervised and self-supervised learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Accelerated MRI acquisition is crucial for reducing scan times.
• Reconstruction methods often struggle with variations in signal-to-noise ratio (SNR) and limited fully sampled data.
• Deep learning models show promise but require extensive labeled data for optimal performance.

Purpose of the Study:

• To develop a novel deep learning method, Noise2Recon, for accelerated MRI reconstruction.
• To enhance robustness against signal-to-noise ratio (SNR) variations and distribution shifts.
• To enable training with limited fully sampled (labeled) and abundant undersampled (unlabeled) MRI data.

Main Methods:

• Proposed Noise2Recon, a consistency training approach for self-supervised learning.
• Utilized both fully sampled and undersampled scans, enforcing reconstruction consistency between undersampled and noise-augmented scans.
• Compared Noise2Recon against compressed sensing and supervised/self-supervised deep learning baselines on knee and brain MRI datasets.

Main Results:

• Noise2Recon achieved superior performance in structural similarity, peak signal-to-noise ratio, and normalized-RMS error in label-limited settings.
• Outperformed all baselines in low-SNR conditions and when generalizing to out-of-distribution (OOD) acceleration factors.
• Demonstrated comparable performance to fully supervised models trained with significantly more data.

Conclusions:

• Noise2Recon offers a label-efficient and robust solution for accelerated MRI reconstruction.
• The method effectively handles distribution shifts, including SNR changes and varying acceleration factors.
• Noise2Recon shows potential for improving MRI reconstruction quality with limited training data.