FDuDoCLNet: Fully dual-domain contrastive learning network for parallel MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) is crucial for soft tissue imaging but suffers from slow acquisition speeds, necessitating faster reconstruction techniques.
• Deep learning (DL) methods have shown promise in MRI reconstruction but often overlook multi-domain features or overemphasize single domains, leading to information loss.
• Existing dual-domain approaches may struggle to balance global structures and local details in reconstructed MR images.

Purpose of the Study:

• To develop an accelerated MRI reconstruction method that effectively utilizes information from both the image and wavelet domains.
• To address limitations of current DL-based MRI reconstruction, particularly the neglect of diverse frequency features and the imbalance in dual-domain analysis.
• To introduce a novel network architecture inspired by wavelet theory's lifting scheme for enhanced reconstruction performance.

Main Methods:

• Proposed a Fully Dual-Domain Contrastive Learning Network (FDuDoCLNet) based on variational networks (VarNet).
• Integrated cascaded dual-domain regularization units and data consistency (DC) layers within the network architecture.
• Introduced a novel dual-domain contrastive loss function to optimize reconstruction by considering both image and wavelet domain features.
• Leveraged the lifting scheme from wavelet theory to enhance feature extraction and representation.

Main Results:

• The FDuDoCLNet achieved a Peak Signal-to-Noise Ratio (PSNR) of 34.439 dB and a Structural Similarity Index Measure (SSIM) of 0.895 on the fastMRI multi-coil knee dataset.
• Demonstrated effective acceleration of MRI reconstruction under a 6× acceleration factor.
• Showcased the network's ability to preserve both global structures and fine details by analyzing features across dual domains.

Conclusions:

• The proposed FDuDoCLNet effectively accelerates MRI reconstruction while maintaining high image quality.
• Dual-domain analysis, particularly with contrastive learning, offers a significant advantage over single-domain or less integrated dual-domain methods.
• The FDuDoCLNet represents a promising advancement in deep learning for medical image reconstruction, particularly for accelerating MRI scans.