Figure 3.

Physics-informed deep learning framework for DECT parametric mapping where the orange and blue arrows denote training and application workflows. The orange arrows represent the training workflow that starts from using DECT images of the electron density phantom as inputs for ML/DL models. The physics model in (g) is only for physics-informed training instead of conventional training. The blue arrows denote the application workflow that starts from using DECT images of CIRS anthropomorphic phantoms as testing data for ML/DL models. The application workflow outputs RSP and mass density (${\mathit{\rho }}_{\mathit{m}}$) maps. The $\text{σ}$ and LN denote activation functions of ReLU (Nair and Hinton, 2010) and layer normalization (Jimmy Lei Ba, 2016). ConvA and ConB are two different convolutional layers defined in Table A1 (Appendix A). The RN block is described in Table A1.