CSSL-ISRVN: consistency self-supervised learning integrating ISTANet and sensitivity refinement-enhanced variational...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) is crucial for its non-invasive nature and soft-tissue contrast.
• Long MRI acquisition times limit clinical use, causing artifacts and patient discomfort.
• Current deep learning methods often require fully sampled data, which is difficult to obtain.

Purpose of the Study:

• To develop a self-supervised framework for MRI reconstruction that does not rely on fully sampled training data.
• To accelerate MRI acquisition times while maintaining high image quality.
• To overcome limitations of existing deep learning approaches in MRI reconstruction.

Main Methods:

• Proposed CSSL-ISRVN, a consistency self-supervised reconstruction framework utilizing an Improved Variational Network (IVN) and Sensitivity Refinement Module (SRM).
• The IVN integrates a Feature Refinement and Denoising Module (FRDM) with residual blocks and a Gaussian Context Transformer for feature extraction.
• SRM refines sensitivity modulation in a closed loop, adaptively adjusting the forward model and data consistency to minimize errors from fixed sensitivity estimates.
• Developed ISRVN, a cascade of ISTANet and SRVN, for artifact suppression and physics-consistent refinement.
• Implemented a consistency self-supervised scheme using calibration and consistency losses to train networks without fully sampled data.

Main Results:

• CSSL-ISRVN demonstrated superior performance compared to existing self-supervised and scan-specific methods on three public datasets, especially with 1D undersampling.
• The framework achieved reconstruction quality competitive with state-of-the-art supervised models.
• Results indicate robustness and stability in MRI reconstruction.

Conclusions:

• CSSL-ISRVN provides an effective solution for accelerated MRI reconstruction, eliminating the need for fully sampled training data.
• The framework's novel integration of sensitivity refinement, hybrid modeling, and self-supervision ensures robust, high-fidelity reconstructions.
• CSSL-ISRVN shows significant potential for practical application in clinical settings, enhancing MRI efficiency and patient experience.