Abstract

Current model-based variational methods used for solving the nonlinear material decomposition problem in spectral computed tomography rely on prior knowledge of the scanner energy response, but this is generally unknown or spatially varying. We propose a twostep deep transfer learning approach that can learn the energy response of the scanner and its variation across the detector pixels. First, we pretrain U-Net on a large data set assuming ideal data, and, second, we fine-tune the pretrained model using few data corresponding to a non-ideal scenario. We assess it on numerical thorax phantoms that comprise soft tissue, bone and kidneys marked with gadolinium, which are built from the kits19 dataset. We find that the proposed method solves the material decomposition problem without prior knowledge of the scanner energy response. We compare our approach to a regularized Gauss-Newton method and obtain a superior image quality.