WUTrans: Whole-spectrum unilateral-query-secured transformer for 4D CBCT reconstruction

Four-dimensional cone-beam computed tomography (4D-CBCT) imaging technology provides images with high spatial and temporal resolution for intraoperative guidance, facilitating real-time tracking of tumor position changes during radiotherapy. However, due to the cyclic motion of respiration causing insufficient projections in each respiratory phase, 4D-CBCT reconstruction suffers from severe streak artifacts. Direct reconstruction from sparsely sampled projection data faces dilemmas between spatial resolution, temporal resolution, and image quality. Unlike the in-room CBCT images, earlier obtained planning CT (pCT) images of the same patient are artifact-free and have the potential to improve 4D-CBCT image qualities. Therefore, we propose a Whole-spectrum Unilateral-query-secured Transformer for Motion Compensation reconstruction (WUTrans-MoCo) with guidance from pCT. First, the Neural Radiance Field (NeRF)-guided joint-optimization network (NGJO-Net) is proposed to fully leverage prior knowledge under the guidance of NeRF for reducing streak artifacts in low-dose sparse-view (SV) reconstruction. Second, to maintain respiratory motion while globally suppressing artifacts, a Whole-spectrum Unilateral-query-secured Transformer Network (WUTran) is constructed for motion compensation on the SV reconstruction results. Third, to further reduce the local anatomical structure differences between pCT and intraoperative CBCT, Structure-Enhanced Feature Fusion (SEFF) is designed for image fusion of the motion-compensated results. The method was evaluated on two public lung CT/CBCT datasets with tumors. Qualitative and quantitative results indicate that WUTrans-MoCo has the potential to reconstruct high-quality 4D-CBCT images with respiratory motion, thereby enhancing the accuracy of radiotherapy and reducing surgical time.