Material Decomposition in Spectral CT Using Deep Learning: A Sim2Real Transfer Approach

The state-of-the art for solving the nonlinear material decomposition problem in spectral computed tomography is based on variational methods, but these are computationally slow and critically depend particular choice of regularization functional. Convolutional neural networks have been proposed addressing issues. However, learning algorithms require large amounts experimental data sets. We propose a deep strategy U-Net architecture Sim2Real transfer approach where knowledge that we learn from synthetic transferred to real-world scenario. In order this work, must be realistic representative data. For purpose, numerical phantoms generated human CT volumes KiTS19 Challenge dataset, segmented into specific materials (soft tissue bone). These projected sinogram space simulate photon counting data, taking account energy response scanner. compared projection- image-based approaches network trained decompose either projection or image domain. strategies regularized Gauss-Newton (RGN) method phantom thorax