Measuring Tissue Elastic Properties Using Physics Based Neural Networks

Ultrasound elastography is a non-invasive and low-cost imaging technique that is used to detect abnormalities in soft tissues. Elastography detects solid tumors from healthy tissues by observing changes in elasticity of tissues on application of force. Reconstruction of initial tissue modulus distribution based on measured displacement/strain fields is called an inverse elasticity problem which has a wide range of applications in medical diagnosis. This paper tries to measure the elastic properties of tissues using Physics-Informed Neural Networks (PINNs). The input data consists of pre-compression and post-compression images of a phantom. Displacement and strain fields are computed from input data which are fed to our PINN model. The PINN model consists of five independent feed-forward neural networks. The model is trained using a loss function that incorporates physics laws based on linear elasticity along with the input data. Lame constants ($\lambda$ and $\mu$) are considered as network parameters that change during the training phase. The ground truth $\lambda$ value is 920 kPa whereas the value predicted by the model is 925.319 kPa. The results indicated that that PINNs can solve inverse problems in the domain of ultrasound elastography.