A Deep-Learning-Based Method for Correction of Bone-Induced CT Beam-Hardening Artifacts

The X-ray attenuation coefficients generally decrease as the X-ray energy increases, which leads to beam-hardening artifacts in computed tomography (CT). Due to the difference of dependence of the attenuation coefficients on energy for soft tissue and bone in human body, a simple water precorrection procedure was unable to correct the bone-induced artifacts. Conventional empirical beam-hardening correction (EBHC) method rely on empirical image segmentation and data combination processes and may not be able to fully correct the artifacts. We developed a physics-driven deep-learning-based method, which followed the workflow of the EBHC method but replaced the empirical components of the EBHC method with neural networks. Numerical experiments were performed to validate the proposed method and benchmark its performance with the EBHC method and the end-to-end training strategies based on two popular neural networks, i.e., U-net and RED-CNN. Results demonstrate that the proposed method achieved the best performance in both qualitative and quantitative aspects.