Dynamic Representation Learning with Spatial-Frequency Features for Motion Compensated 4D Cone-Beam Reconstruction

Four-dimensional cone-beam computed tomography (4D CBCT) is a promising tool to overcome the respiratory induced organ motion and alleviate the serious negative effects on radiation therapy treatment efficiency. However, the projections collected from a single CBCT scan is extremely sparse for the reconstruction of the ten respiratory phases, and result in the reconstruction results with serious streak artifacts and lacking of the vital organ motion information. Thus, a dynamic representation learning framework with spatial-frequency features is proposed in this study, which consists of a spatial-frequency representation enhancement network and a spatial-frequency motion representation network. Towards the streak artifacts caused by the sparsely sampled projections within each phase, spatial-frequency representation enhancement network is presented to capture both the global and local features as well as their relations by spatial-Fourier blocks to restore the anatomical details as possible. Based on the enhanced phase-correlated images, spatial-frequency motion representation network is proposed to obtain the DVFs among various respiratory phases and achieve the precise motion compensated 4D CBCT reconstruction. Both of the simulated and clinical 4D CBCT datasets were performed to evaluate the performance of the proposed framework. Compared with the existing methods, the proposed reconstruction method achieves promising results in streak artifact reduction, anatomical detail restoration, and organ motion representation.