Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Deep learning models for MRI reconstruction typically require fully sampled datasets for training.
• Acquiring fully sampled data is time-consuming and limits acceleration in MRI scans.
• Developing training strategies that do not rely on fully sampled data is crucial for advancing MRI technology.

Purpose of the Study:

• To develop a strategy for training physics-guided MRI reconstruction neural networks without requiring fully sampled data.
• To enable self-supervised training of deep learning models for MRI reconstruction using undersampled measurements.

Main Methods:

• Implemented a self-supervised learning via data undersampling (SSDU) approach for physics-guided deep learning MRI reconstruction.
• Partitioned available measurements into two sets for data consistency units and loss definition within an unrolled network.
• Compared SSDU training with fully supervised training and conventional compressed-sensing and parallel imaging methods on the fastMRI knee database.

Main Results:

• The SSDU approach achieved performance comparable to supervised learning on knee MRI data at 4x acceleration.
• SSDU significantly outperformed conventional compressed-sensing and parallel imaging methods based on quantitative metrics and a clinical reader study.
• The method successfully reconstructed prospectively undersampled brain datasets at high acceleration rates (4x, 6x, 8x) where supervised learning is not feasible.

Conclusions:

• Self-supervised learning via data undersampling (SSDU) enables effective training of physics-guided deep learning MRI reconstruction without fully sampled data.
• The proposed SSDU approach achieves comparable results to supervised deep learning MRI reconstruction.
• SSDU offers a viable solution for training advanced MRI reconstruction models, particularly in scenarios lacking ground-truth data.