Search

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

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