Self-supervised learning for MRI reconstruction through mapping resampled k-space data to resampled k-space data.

Area of Science:

• Medical Imaging
• Deep Learning
• Magnetic Resonance Imaging

Background:

• Traditional MRI reconstruction requires fully sampled data, which is often difficult to acquire in clinical settings due to physiological and physical constraints.
• Deep learning (DL) has shown promise in MRI reconstruction, but often still relies on fully sampled data for training.

Purpose of the Study:

• To introduce a novel self-supervised deep learning approach for accurate and efficient reconstruction of undersampled MRI data.
• To develop a method that does not require fully sampled reference datasets for training, addressing limitations in real-world clinical scenarios.

Main Methods:

• The proposed method, Randomly Self-Supervised Learning via Data Undersampling (RSSDU), resamples k-space data twice to create two subsets with varying acceleration factors.
• A network is trained in a supervised manner to learn the mapping between these two undersampled data subsets.

Main Results:

• RSSDU demonstrated superior performance compared to established self-supervised methods like SSDU and K-band.
• Performance was evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index measurement (SSIM).

Conclusions:

• RSSDU offers an efficient and accurate solution for reconstructing MRI images from undersampled data without needing fully sampled references.
• The method shows significant potential for improving MRI acquisition efficiency and applicability in clinical practice.