UNCERTAINTY-GUIDED PHYSICS-DRIVEN DEEP LEARNING RECONSTRUCTION VIA CYCLIC MEASUREMENT CONSISTENCY.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Science

Background:

• Physics-driven deep learning (PD-DL) enhances computational imaging, particularly in MRI.
• PD-DL integrates physical models for data fidelity and uses neural networks for regularization.
• Current PD-DL training progresses from supervised to self-supervised and generative models, with uncertainty quantification efforts focused on generative models.

Purpose of the Study:

• To develop an uncertainty estimation process for the data fidelity component of PD-DL.
• To guide the training of PD-DL methods using the derived uncertainty estimates.
• To improve the reconstruction quality of PD-DL methods in MRI applications.

Main Methods:

• Devised an uncertainty estimation process focusing on PD-DL's data fidelity component.
• Characterized uncertainty by analyzing cyclic consistency between different forward models.
• Utilized the uncertainty estimate to guide the PD-DL training procedure.

Main Results:

• The proposed uncertainty estimation effectively characterized the data fidelity component.
• Uncertainty-guided training led to improved reconstruction quality in PD-DL methods.
• Demonstrated the efficacy of cyclic consistency for uncertainty quantification in PD-DL.

Conclusions:

• Uncertainty estimation based on cyclic consistency is a valuable tool for PD-DL.
• Guiding PD-DL training with uncertainty improves overall reconstruction performance.
• This approach offers a novel strategy for enhancing PD-DL in medical imaging.