Abstract
BACKGROUND: Cone-beam CT (CBCT) captures on-board volumetric anatomy for image guidance and treatment adaptation in radiotherapy. To compensate for respiration-induced anatomical motion, time-resolved CBCT is highly desired to capture the spatiotemporal anatomical variations but faces challenges in accuracy and efficiency due to substantial optimization needed in image reconstruction and motion modeling. PURPOSE: We proposed a fast time-resolved CBCT reconstruction framework, based on a dynamic reconstruction and motion estimation method with new reconstructions initialized and conditioned on prior reconstructions in an adaptive fashion (DREME-adapt). MATERIALS AND METHODS: DREME-adapt reconstructs a time-resolved CBCT sequence from a fractional standard CBCT scan while simultaneously generating a machine learning-based motion model that allows single-projection-driven intra-treatment CBCT estimation and motion tracking. Via DREME-adapt, a virtual fraction is generated from a pre-treatment 4D-CT set of each patient for a clean, "cold-start" reconstruction. For subsequent fractions of the same patient, DREME-adapt uses pre-derived motion models and reference CBCTs as initializations to drive a "warm-start" reconstruction, based on a lower-cost refining strategy. Three strategies: DREME-cs which drops the "warm-start" component, DREME-adapt-vfx which uses a fixed initialization (virtual fraction's reconstruction results), and DREME-adapt-pro which initialize reconstructions through a progressive daisy chain scheme (virtual fraction for fraction 1, fraction 1 for fraction 2, and so on), were evaluated on a digital phantom study (7 motion/anatomical scenarios) and a patient study (seven patients). RESULTS: DREME-adapt allows fast and accurate time-resolved CBCT reconstruction. For the XCAT simulation study, DREME-adapt-pro achieves image reconstruction relative error of 0.14 ± 0.01 and tumor center-of-mass tracking error of 0.92 ± 0.62 mm (mean ± s.d.), compared to 0.15 ± 0.01 and 1.06 ± 0.73 mm for DREME-adapt-vfx, and 0.18 ± 0.01 and 1.96 ± 1.35 mm for DREME-cs. For the real-time motion inference test dataset of the patient study, DREME-adapt-pro localizes moving lung landmarks to a mean ± s.d. error of 2.21 ± 1.79 mm. In comparison, the corresponding values for DREME-adapt-vfx and DREME-cs were 2.53 ± 1.93 mm and 3.22 ± 2.88 mm, respectively. The DREME-adapt-pro training takes 11 min, only 15% of the original DREME algorithm. CONCLUSIONS: With high efficiency and accuracy, DREME-adapt-pro allows on-board time-resolved CBCT reconstruction and enhances the clinical adoption potential of the DREME framework.