Accelerating multi-coil MR image reconstruction using weak supervision.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) reconstruction often requires large datasets, which are difficult to acquire.
• Deep learning methods offer potential for faster and more robust MRI reconstruction.
• Integrating physics-guided principles can improve the accuracy of deep learning models.

Purpose of the Study:

• To evaluate a weakly supervised, multi-coil, physics-guided deep learning approach for MR image reconstruction.
• To leverage both under-sampled and fully sampled datasets for improved reconstruction quality and robustness.
• To enhance MRI acquisition speed and reduce data requirements.

Main Methods:

• A physics-guided variational network (VarNet) was pre-trained using self-supervised learning via data undersampling (SSDU) on under-sampled data.
• Pre-trained weights were transferred and fine-tuned on a smaller, fully sampled dataset using multi-scale structural similarity (MS-SSIM) loss.
• The methodology was compared against fully self-supervised and fully supervised training approaches.

Main Results:

• Demonstrated improved reconstruction quality (SSIM, PSNR, NRMSE) in the high-data regime.
• Showcased enhanced robustness in the low-data regime, crucial for scarce data scenarios.
• Achieved high acceleration rates (8x for knee, 10x for brain MR imaging) with improved performance.

Conclusions:

• Weakly supervised, physics-guided MR image reconstruction using transfer learning is feasible and effective.
• The proposed method offers superior reconstruction quality and robustness compared to traditional methods.
• This approach holds promise for accelerating MRI scans and improving diagnostic accuracy, particularly in data-limited situations.