RSME: Respiration-Driven Synchronized Motion Estimator for Real-Time Thoracic 3-D CT Reconstruction Using Low-Rank Motion Fields and Dose-Free Surface Imaging

Real-time 3-D CT reconstruction during radiotherapy, based on planning 4-D CT, has emerged as a central area of interest in the field of image-guided radiotherapy (IGRT). However, current methodologies still rely on time-consuming patient-specific training, real-time registration, or additional radiation doses. Our contribution, respiration-driven synchronized motion estimator (RSME), represents a novel and efficient forward propagation network, leveraging an explicit spatial-temporal low-rank decomposition of displacement fields and surface imaging to enable real-time 3-D CT reconstruction during respiration. RSME is derived from a well-designed, interpretable inverse optimization problem, mapping static spatial components and dynamic skin depth images to dynamic temporal components for reconstruction. In RSME, we customize two core transformer-based encoders: the motion former (M-Former) and the motion pattern former (P-Former), and incorporate customized cross attention mechanism to effectively gauge the interdependencies between the two encoders. Strategically, we input skin depth images into M-Former to inquire about motion information, thus circumventing the requirement for real-time registration. In P-Former, we introduce static spatial components that integrate explicit respiratory pattern details to distinguish between patients without the necessity for patient-specific training, and harness the static nature to evade additional inference latency. Extensive experimental results demonstrate that RSME achieves comparable or even superior accuracy than state-of-the-art methods, with a cumulative latency of 96 ms. Notably, RSME accomplishes this without necessitating additional radiation dosage, time-consuming patient-specific training or real-time registration.