Robust multimaterial decomposition of spectral CT using convolutional neural networks

Spectral computed tomography (CT) can reconstruct scanned objects at different energy-bins and thus solve the multimaterial decomposition (MMD) problem. Because the linear attenuation coefficients of different basis materials may be extremely close, the decomposition problem is often ill-conditioned. Meanwhile, traditional material decompositions with image-domain algorithms are usually voxelwise based. Therefore, these algorithms rely heavily on image quality. Ring artifacts often exist in the reconstructed images of spectral CT due to the inconsistency feature of energy-resolved detectors and beam-hardening effect. Considering the enlargement of the receptive field and taking advantage of the modeling ability of convolutional neural networks in deep learning, we proposed a convolutional material decomposition algorithm to solve the MMD problem through a basis of patches instead of pixels of the spectral CT images. Simulations and physical experiments were performed to validate the proposed algorithm, and its quality was compared with a traditional MMD algorithm in the image domain. Results show that the proposed method achieves good accuracy, reduces mean squared errors by one to two orders, and exhibits robustness in the MMD of spectral CT images even in the case that obvious ring artifacts is presented.