A Deep Learning Normalization Method for Robust Ultrasound Attenuation Coefficient Estimation

The Ultrasonic Attenuation Coefficient (AC) is a pivotal biomarker in clinical diagnostics, instrumental in assessing conditions like non-alcoholic fatty liver disease and distinguishing tumors. Traditional methods, including the frequency shift method and reference phantom techniques, often fall short— the former due to instability in AC estimation and the latter due to impractical calibration requirements. Addressing these limitations, our study proposes a deep learning-based method for the time-frequency domain normalization of ultrasound signals. The proposed method using convolutional neural network (CNN) aims to decouple the ultrasound imaging system parameters, the backscatter coefficient, and the tissue acoustic parameters to obtain an accurate estimate of the AC. The CNN filters out system parameter and backscatter coefficient influences from the RF signal spectra, enhancing the precision and reliability of AC estimations. Rigorous validation using simulated and phantom ultrasound data corroborates the superior accuracy and robustness of our proposed technique in determining the AC.