Construction of Simulated 3D CT from Multiple-View X-ray Images

Computed tomography (CT) images provide reliable information for clinical diagnosis and navigation during surgery. However, because the acquisition of traditional CT during surgery remains challenging, preoperative CT images are usually used for surgical navigation in clinical settings. With such use of previously acquired images, body movement, organ deformation, and surgical procedures may cause changes in the locations of lesions, making pre-generated CT data incapable of providing accurate information. To address this issue, the reconstruction of 3D CT data from intra-operative 2D X-ray images is a promising option. Although previous work reconstructed CT from biplanar X-ray images using a generative adversarial network, the reconstruction quality was far from that required in clinical settings because of a lack of 3D topology information. In this work, we investigate how many views are necessary to reconstruct adequate CT images, and introduce a novel architecture called Muti-view X-ray Aware GAN(MXA-GAN) for 2D-to-3D image reconstruction. This architecture involves utilizing rotations and concatenations from X-ray images acquired with different views. We experimentally validated our model on the publicly available chest CT dataset LIDC-IDRI, demonstrating the effectiveness of our proposed approach.