FCSSL: fusion enhanced contrastive self-supervised learning method for parallel MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning significantly accelerates magnetic resonance imaging (MRI) but requires fully sampled datasets, which are often unfeasible or costly to acquire.
• Existing methods struggle with data limitations, hindering widespread application in resource-constrained scenarios.

Purpose of the Study:

• To develop a novel self-supervised learning method for parallel MRI reconstruction that eliminates the need for fully sampled k-space data and coil sensitivity maps.
• To enhance the representational capacity and reconstruction quality through a contrastive learning framework and an adaptive fusion network.

Main Methods:

• Proposed a fusion enhanced contrastive self-supervised learning (FCSSL) method for parallel MRI reconstruction.
• Implemented a contrastive learning framework with re-undersampling masks to improve representational capacity.
• Designed a novel adaptive fusion network trained in a self-supervised manner to integrate reconstruction results.

Main Results:

• FCSSL demonstrated superior reconstruction performance compared to other self-supervised methods on knee datasets.
• FCSSL performance approached supervised methods without requiring fully sampled data, particularly under 2DRU and RADU masks.
• The model showed effective generalization to unseen undersampling masks and achieved comparable performance after minimal fine-tuning on new data.

Conclusions:

• FCSSL provides a viable solution for high-quality MRI reconstruction without fully sampled datasets.
• The method overcomes significant hurdles in scenarios where acquiring complete MR data is challenging.
• This approach advances self-supervised learning applications in medical imaging, enabling faster and more accessible MRI.