vSHARP: Variable Splitting Half-quadratic ADMM algorithm for reconstruction of inverse-problems.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Medical Imaging (MI) reconstruction often involves solving ill-posed inverse problems.
• Traditional methods like Compressed Sensing (CS) can be slow and yield low-fidelity images.
• Deep Learning (DL) methods show promise in surpassing conventional approaches for inverse problems.

Purpose of the Study:

• To introduce vSHARP, a novel DL-based method for reconstructing images from noisy or incomplete measurements in MI.
• To address the limitations of existing methods in solving ill-posed inverse problems.
• To enhance image quality and reconstruction speed in accelerated parallel MRI.

Main Methods:

• vSHARP employs Half-Quadratic Variable Splitting and ADMM for optimization.
• A differentiable gradient descent process ensures data consistency in the image domain.
• A DL-based denoiser (e.g., U-Net) and a dilated-convolution model are utilized for image enhancement and parameter prediction.

Main Results:

• vSHARP demonstrates superior performance in accelerated parallel MRI reconstruction.
• Evaluated on two distinct datasets for static MRI and one for dynamic MRI.
• Comparative analysis shows improved results over state-of-the-art methods.

Conclusions:

• vSHARP offers a powerful DL-based solution for ill-posed inverse problems in medical imaging.
• The method achieves high-fidelity image reconstruction in accelerated MRI tasks.
• vSHARP represents a significant advancement in medical image reconstruction technology.