aDiner: Adaptive Dynamic Implicit Neural Representation for Dynamic CBCT Imaging

Cone-beam Computed Tomography (CBCT) is essential for target localization and treatment planning in image-guided radiotherapy for lung cancer. Dynamic reconstruction is an ill-posed inverse problem, as each motion state is captured by only one single projection. Several supervised learning methods have been proposed, but they rely on training with paired simulated or real datasets, and reconstruct images at discrete respiratory phases. Therefore, we propose a self-supervised learning method for lung imaging in the radiotherapy, adaptive Dynamic implicit neural representation (aDiner), to reconstruct dynamic CBCT images for each timepoint. Specifically, an adaptive composite representation (ACR) is introduced. In ACR, the inter- and intra-cycle encoding (iICE) captures the inter- and intra-cycle similarities of lung motion and ensures temporal coherence. Moreover, a probability-based feature fusion (PFF) is proposed to fuse decoupled static and dynamic features, composing the spatiotemporal representation. To improve the motion estimation, we further design a motion-guided coarse-to-fine sampling (MCS) strategy to focus on dynamic regions, with adaptive projection-ray sampling (APS) and adaptive deformed-point sampling (ADS). Compared with the previous methods, aDiner separately estimates static and dynamic regions, and takes advantage of the inter- and intra-cycle similarities to achieve better reconstruction quality. aDiner was evaluated on the simulated dataset, the physical phantom, and the clinical dataset. The results showed that aDiner could accurately and robustly reconstruct dynamic CBCTs for each timepoint. It also captured the lung motion in both regular and irregular respirations. The dynamic demos are available at: https://github.com/Henryzyf/aDiner.