PEARL: Cascaded Self-Supervised Cross-Fusion Learning for Parallel MRI Acceleration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Supervised deep learning (SDL) for accelerated magnetic resonance imaging (AMRI) requires extensive training data.
• Self-supervised methods like Deep Image Prior (DIP) avoid explicit training but struggle with noise and artifacts in highly degraded images.

Purpose of the Study:

• To introduce PEARL, a novel self-supervised framework for accelerated parallel MRI.
• To enable accurate image reconstruction from compressively sampled k-space data without extensive training datasets.

Main Methods:

• PEARL utilizes a multiple-stream joint deep decoder with cross-fusion schemes.
• Each stream employs cascaded sub-block networks (SBNs) with combined upsampling, 2D convolution, joint attention, and batch normalization.
• Dual-normalized edge-orientation similarity regularization is incorporated into the loss function for enhanced reconstruction and overfitting prevention.

Main Results:

• PEARL demonstrates superior performance compared to state-of-the-art self-supervised AMRI methods.
• Accelerated acquisitions (5-6x) showed significant improvements: 1-2% in SSIMROI, 3-6% in PSNRROI, and a 15-20% reduction in RLNEROI.

Conclusions:

• PEARL offers an effective self-supervised solution for accelerated MRI, overcoming limitations of data dependency and image degradation.
• The proposed architecture and regularization enhance image quality and reconstruction accuracy in AMRI.