Deep Learning for Ultrasound Attenuation Coefficient Estimation

Abstract

Estimating the attenuation coefficient slope (ACS) in quantitative ultrasound (QUS) plays an important role in providing objective diagnostic information about tissue characteristics. Different methods, including the spectral log difference (SLD) or the regularized spectral log difference (RSLD), have been used to estimate ACS, but they face limitations, such as the need to balance spatial resolution and accuracy or the requirement for manual tuning of regularization parameters prior to the estimation. This study investigates the use of a Deep Neural Network (DNN) with U-Net architecture for estimating ACS maps, comparing its performance against the RSLD technique. The DNN was trained on simulated data generated with Kwave. Evaluation involved both simulated and experimental data acquired from CIRS phantoms. Results show that the DNN outperforms RSLD, achieving mean MAPE, SDPE, and CNR of 1.25, 1.88, and 36.3 dB in simulations, respectively, compared to 10.42, 12.22, and 3.42 dB for RSLD. Experimental validation confirmed superior DNN performance with MAPE, SDPE, and CNR values of 8.25, 12.72, and 3.92 for Phantom 1, and 5.83, 8.06, and 1.85 for Phantom 2. The DNN demonstrated better shape capture and accuracy in ACS estimations without the need for manual regularization tuning, indicating its robustness and capability to generalize from simulations to experimental data.