On Noise-Robust Deep Learning-Based Viscoelasticity Modeling for Ultrasound Shear Wave Elastography

Abstract

This thesis investigates noise-robust viscoelasticity modeling for ultrasound shear wave elastography, with a focus on developing deep learning and data-driven approaches for reliable characterization of tissue mechanical properties. The work addresses key challenges associated with low signal-to-noise ratio, tissue heterogeneity, and the restrictive assumptions of conventional model-based inversion methods. The thesis first develops a deep learning–based denoising framework operating in the time–frequency domain to suppress noise in shear wave velocity fields. The proposed approach improves signal quality and enables robust reconstruction of shear wave phase velocity maps under highly noisy conditions. Its performance is evaluated using simulated data, tissue-mimicking phantoms, and ex vivo biological tissues. Building on this improved signal fidelity, a unified multitask deep learning framework is then proposed for the simultaneous reconstruction of elasticity and viscosity maps directly from noisy shear wave data. By incorporating a dedicated denoising module, the approach reduces reliance on explicit shear wave attenuation estimation and restrictive physical assumptions. The framework demonstrates accurate and robust reconstruction across numerical simulations, tissue-mimicking phantoms, and in vivo animal liver data, including challenging low-SNR conditions. The thesis further investigates the limitations of conventional shear wave dispersion models and proposes an empirically modified power-law formulation to capture nonlinear frequency-dependent behavior over an extended bandwidth. Overall, the thesis demonstrates that noise-aware deep learning and data-driven viscoelasticity modeling can substantially improve the accuracy, robustness, and diagnostic potential of ultrasound shear wave elastography, particularly under practical imaging conditions characterized by low signal quality.