PARCEL: Physics-Based Unsupervised Contrastive Representation Learning for Multi-Coil MR Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Deep learning has advanced magnetic resonance (MR) imaging, particularly for parallel imaging techniques.
• Current neural network-based methods face limitations due to the lack of high-quality, fully sampled training datasets and insufficient model interpretability.

Purpose of the Study:

• To propose a novel method, Physics-bAsed unsupeRvised Contrastive rEpresentation Learning (PARCEL), to accelerate parallel MR imaging.
• To address the limitations of existing methods by enabling accurate MR reconstruction without fully sampled datasets and enhancing model interpretability.

Main Methods:

• PARCEL employs a parallel framework with two branches of model-based unrolling networks.
• It utilizes augmented undersampled multi-coil k-space data for contrastive representation learning.
• A sophisticated co-training loss guides the networks to capture inherent MR image features and representations.

Main Results:

• PARCEL demonstrated effective learning of essential representations for accurate MR reconstruction.
• Evaluation on two in vivo datasets showed competitive performance compared to five state-of-the-art methods.
• The method successfully reconstructs MR images without reliance on fully sampled datasets.

Conclusions:

• PARCEL offers a robust solution for accelerating parallel MR imaging by leveraging unsupervised contrastive learning.
• The proposed method overcomes the dependency on fully sampled datasets, enhancing the practicality of deep learning in MR imaging.
• PARCEL provides a promising direction for developing more interpretable and efficient MR reconstruction techniques.