Dual-path deep inverse learning of piezoelectric shear-wave responses for spatial viscoelastic modulus reconstruction and mechanical characterization of endometrial tumors

Endometrial tumors exhibit heterogeneous mechanical properties that may provide valuable information for tumor characterization, while conventional shear-wave elastography often provides limited spatial representation of viscoelastic behavior. This study presents a Dual-Path Deep Inverse Learning model to recover spatially varying viscoelastic properties from piezoelectric shear-wave responses. The aim was to learn the complementary spatial and temporal wave properties independently, combine these representations, and reconstruct the elastic modulus and viscosity by physics-guided inverse learning. The computational data set of 300 virtual endometrial tumor models was increased to 6,000 shear-wave samples by varying tumor geometry, mechanical properties, excitation conditions, and noise. The data was split into training, validation and test sets. The proposed model integrated feature extraction in both space and time, feature fusion, inverse reconstruction, physics-guided learning, and spatial mechanical mapping. The framework obtained an MAE of 0.072, an RMSE of 0.096 and a correlation of 0.970. The differences in reconstruction of the elastic modulus and the viscosity were 2.39% and 2.16%, respectively. Mechanical characterization, on the other hand, had a 95.8% accuracy and an AUC of 0.976. The results showed accurate spatial viscoelastic reconstruction, strong tumor characterization ability, and better results than comparative methods.