Integrating Learning-Based Priors With Physics-Based Models in Ultrasound Elasticity Reconstruction.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Science

Background:

• Ultrasound elastography visualizes tissue stiffness by solving inverse problems.
• Classical methods use regularization, often causing blurriness and artifacts.
• Finding optimal regularizers for elasticity reconstruction is challenging.

Purpose of the Study:

• To develop an improved statistical framework for ultrasound elastography image reconstruction.
• To incorporate signal-dependent noise modeling and learning-based priors.
• To reduce artifacts and enhance the accuracy of elasticity reconstructions.

Main Methods:

• A novel statistical representation of the physical imaging model was developed.
• A learning-based framework integrated physical models with learned priors (denoising regularizer).
• Optimization was performed using fixed-point and gradient descent methods within plug-and-play (PnP) and regularization by denoising (RED) paradigms.

Main Results:

• Achieved nearly 20% improvement in relative mean square error (RMSE) for piecewise smooth phantoms.
• Demonstrated 12% RMSE improvement for breast-mimicking phantoms.
• Qualitative results showed robust performance for complex elasticity structures.

Conclusions:

• The proposed learning-based framework significantly enhances ultrasound elastography reconstruction accuracy.
• The method shows potential for improved clinical diagnosis due to reduced artifacts and better image quality.
• This approach offers a promising alternative to classical model-based methods for quantitative tissue stiffness imaging.