A CONVEX COMPRESSIBILITY-INSPIRED UNSUPERVISED LOSS FUNCTION FOR PHYSICS-DRIVEN DEEP LEARNING RECONSTRUCTION.

Area of Science:

• Medical Imaging
• Machine Learning
• Image Reconstruction

Background:

• Physics-driven deep learning (PD-DL) is increasingly used for fast MRI reconstruction.
• Early PD-DL methods primarily employed supervised learning.
• Recent interest focuses on unsupervised learning for PD-DL training.

Purpose of the Study:

• To propose a novel convex loss function for PD-DL training.
• To evaluate reconstruction quality in supervised, unsupervised, and zero-shot scenarios.
• To offer an alternative learning strategy inspired by compressed sensing.

Main Methods:

• Developed a convex loss function evaluating image compressibility and data fidelity.
• Utilized the reweighted L1 norm as an approximation for the L0 norm.
• Applied the loss function to PD-DL networks for MRI reconstruction.

Main Results:

• PD-DL networks trained with the proposed loss outperformed conventional methods.
• Reconstruction quality was comparable to existing supervised and unsupervised PD-DL techniques.
• The novel loss function demonstrated versatility across different training settings.

Conclusions:

• The proposed convex loss function is an effective alternative for PD-DL training.
• This approach enhances MRI reconstruction quality and generalizability.
• It offers a promising direction for future research in accelerated MRI acquisition.