Material Decomposition Method for Dual-MeV Energy CT via Convolutional Neural Network

With the expansion of the application field of Dual-energy computed tomography(DECT), megavolt(MV) X-rays DECT are gradually needed in some specific scenarios. However, there are truly challenges in the reconstruction as well as the material decomposition of MV DECT, which result in the failure of classic material decomposition methods. In this paper, we propose a CNN-based material decomposition method for MV DECT and prove the feasibility of CNN method in this problem. The main idea is that we take not only the pixel we want to decompose into consideration, but also the patch around the pixel to improve the accuracy of decomposition. As experiment results shown, the CNN method has better performance than classic projection-domain method. The proposed CNN-based method can reduce the MSE of electron density rho(e) by 2 similar to 3 order, and reduce the MSE of atomic number Z by 4 similar to 5 order.