ACS-Net: A Deep Unfolded ADMM Framework for Ultrasound Attenuation Imaging

Abstract

Ultrasound attenuation imaging is gaining traction for its promising clinical diagnostic applications. Estimation methods such as Spatially Weighted Fidelity and Regularization Terms (SWIFT) and its deep learning-aided variant (DL-SWIFT) have demonstrated improved contrast-to-noise ratio (CNR) and more consistent attenuation coefficient slope (ACS) estimates through the use of spatially weighted formulations. However, both methods may still produce artifacts in heterogeneous regions with abrupt backscatter coefficient changes. To address these limitations, we propose ACS-Net, a deep unfolded Alternating Direction Method of Multipliers framework that integrates learned denoising operations within the classical iterative optimization process to reduce ACS estimation bias while preserving inclusion delineation. Phantom experiments confirmed a bias reduction of more than 40% on inclusions, while in vivo thyroid nodule results showed that ACS-Net decreases background coefficient of variation by nearly 50% and improves CNR by more than 30% compared to SWIFT and DL-SWIFT. These findings highlight the clinical promise of deep unfolded optimization methods for reliable and accurate ultrasound attenuation imaging.