PARCEL

PARCEL reconstructs parallel MR images from undersampled multi-coil k-space data using a physics-informed, unsupervised contrastive learning framework that trains two model-based unrolling network branches with a co-training loss to improve reconstruction and interpretability.

Key Features:

• Physics-informed learning: Employs a physics-informed approach to embed MR imaging physical principles into the learning process.
• Unsupervised learning: Uses unsupervised learning to remove dependence on fully sampled datasets.
• Contrastive learning on augmented k-space: Applies contrastive learning to augmented undersampled multi-coil k-space data within a parallel framework.
• Parallel dual-branch unrolling: Implements two branches of model-based unrolling networks trained simultaneously in parallel.
• Co-training loss with three components: Optimizes both branches with a co-training loss composed of three explicit components to capture essential features and representations.
• Interpretability: Integrates physics-based principles to enhance interpretability of the learned models.

Scientific Applications:

• Parallel MR imaging reconstruction: Improves speed and accuracy of parallel MR image reconstruction from undersampled multi-coil k-space.
• Clinical undersampled acquisition: Enables reconstruction workflows that operate without fully sampled datasets, addressing scarcity in clinical acquisitions.
• Evaluation on in vivo datasets: Has been evaluated on two in vivo datasets and demonstrated superior performance compared to five state-of-the-art methods.
• Research on MR reconstruction methods: Facilitates development and comparison of unsupervised, physics-informed MR reconstruction algorithms.

Methodology:

Uses a parallel framework where two model-based unrolling network branches are trained simultaneously via contrastive learning on augmented undersampled multi-coil k-space data, guided by a co-training loss composed of three components.