Author: Hua-Chieh Shao, You Zhang, Ruizhi Zuo 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Cone-beam CT (CBCT) provides on-board patient anatomy for image guidance and treatment adaptation in radiotherapy. However, to compensate for respiration-induced anatomical motion, motion-resolved CBCT reconstruction could be time-consuming due to intensive computation and motion modeling needs. To address this challenge, we proposed a fast motion-resolved CBCT reconstruction framework, based on a dynamic reconstruction and motion estimation method with new reconstructions initialized and conditioned on prior reconstructions in a progressive, adaptive fashion (DREME-adapt).
Methods: DREME-adapt reconstructs a motion-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, we use the first fractional CBCT scan of each patient for a clean, ‘cold-start’ reconstruction. For subsequent fractions of the same patient, DREME-adapt uses the pre-derived motion model and reference CBCT 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-fx1 which uses a fixed initialization (fraction one’s reconstruction results), and DREME-adapt-pro which initializes via a progressive daisy chain scheme, were evaluated on a set of five lung patients, each with 3-5 fractional CBCT scans from the treatment course.
Results: DREME-adapt allows fast and accurate motion-resolved CBCT reconstruction. Compared with reference lung landmark trajectories directly tracked from cone-beam projections, DREME-adapt-pro localizes moving lung landmarks to an average(±s.d.) error of 2.18±1.77 mm. In comparison, the corresponding values for DREME-adapt-fx1 and DREME-cs were 2.58±2.16 mm and 3.18±2.66 mm, respectively. The training of DREME-adapt takes 27 minutes, only 15% of the original DREME algorithm.
Conclusion: DREME-adapt uses the CBCT and motion model information extracted from previous fractions of the same patient to expedite the reconstruction tasks of the subsequent fractions. It allows more efficient on-board reconstruction and enhances the clinical adoption potential of the DREME framework.