Multi-task learning framework for elasticity and viscosity image reconstruction in ultrasound viscoelastography

Soft tissues are viscoelastic in nature, and shear wave viscoelastography has emerged as an effective tool for differentiating healthy and diseased tissues by assessing their viscoelastic properties. This study presents a noise-aware multitask deep learning framework that simultaneously reconstructs tissue elasticity and viscosity maps. The proposed unified model directly estimates viscoelastic parameters from noisy shear-wave data, eliminating the need for explicit shear wave attenuation measurements and improving robustness to noise. The framework was trained on 1,000 velocity field datasets generated from numerical simulations, and validated across numerical simulations, tissue-mimicking phantoms, and in vivo duck liver datasets. Results demonstrated high fidelity in reconstruction accuracy, with strong agreement with ground truth in simulations and phantoms. In simulated phantoms where ground truth was available, the root mean square error for viscosity across samples was 0.128 (95% CI: 0.105–0.150) and for elasticity was 2.363 (95% CI: 2.157–2.573), and the structural similarity index was more than 0.96. In four in vivo duck liver datasets, elasticity values ranged from 11.50 ± 6.32 to 25.50 ± 17.16 kPa and viscosity ranged from 0.85 ± 0.46 to 1.8 ± 0.37 Pa·s. The denoising module improved the signal-to-noise ratio up to threefold and enabled reliable performance even in low-SNR conditions. The framework enables rapid, noninvasive viscoelastic imaging and establishes a foundation for future diagnostic tools that go beyond stiffness-based assessments.