Integrating Deep Learning into PnP-ADMM for Ultrasound Attenuation Coefficient Estimation

Abstract

Estimating the attenuation coefficient slope (ACS) is essential for tissue characterization in quantitative ultrasound (QUS). Traditional model-based methods such as the regularized spectral log difference (RSLD) rely on manually tuned priors, while recent end-to-end deep learning approaches struggle to generalize to in vivo data. This work proposes a hybrid method that integrates pre-trained Attention U-Nets within a Plug-and-Play ADMM framework. The fidelity term is replaced by a network conditioned on spectral inputs and the iteration index, while a second network acts as a learned regularizer. The method was evaluated on physical phantoms and in vivo breast and thyroid acquisitions, after being trained entirely on simulations. Results show improved accuracy and generalization over RSLD and end-to-end baselines, suggesting that embedding a pre-trained deep learning model within a physics-based framework enhances robustness and enables more reliable ACS estimates in real-world scenarios.