Self-Supervised Feature Learning for Cardiac Cine MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning for MRI reconstruction often requires fully-sampled data, which is difficult to acquire due to long scan times and motion.
• Existing fully-sampled datasets may be biased from conventional reconstruction of accelerated data, limiting potential performance.
• Undersampled datasets in clinical practice are underutilized, presenting an opportunity for improved reconstruction methods.

Purpose of the Study:

• To develop a self-supervised framework for MRI reconstruction that overcomes limitations of supervised learning.
• To learn sampling-insensitive features from undersampled MRI data.
• To improve artifact removal and generalization ability in MRI reconstruction.

Main Methods:

• A self-supervised feature extractor was trained on undersampled MRI images to learn robust features.
• These learned features were integrated into a self-supervised reconstruction network (SSFL-Recon).
• The framework was evaluated retrospectively on a 2D cardiac Cine dataset from 91 patients and 38 healthy subjects.

Main Results:

• The SSFL-Recon framework demonstrated superior performance compared to existing self-supervised MRI reconstruction methods.
• Performance was comparable or better than supervised learning methods, even with up to 16x retrospective undersampling.
• The feature learning strategy effectively extracted global representations, aiding artifact removal and enhancing generalization.

Conclusions:

• Self-supervised feature learning offers a promising approach for MRI reconstruction, particularly when fully-sampled data is unavailable.
• SSFL-Recon effectively addresses the challenge of undersampled data, improving image quality and scan efficiency.
• The proposed method shows significant potential for clinical application, enabling faster and more reliable MRI scans.