Room: Track 2
Purpose: Volumetric surrogate-based motion models (SMMs) have been proposed to guide motion management in radiation therapy. SMMs are constructed based on the a priori correlation between an external surrogate and internal motion during CT-simulation. Changes in correlation often result in model breakdown. We present an SMM that exploits additional fluoroscopic images (FLs) to update the correlation. We validated its performance using nine ~30s FL acquisitions from 5 lung-SBRT patients.
Methods: Under a prospective IRB, 4DCT , VisionRT surfaces (VRT), lateral and anterior-posterior FL (used as reference) were collected. Deformation vector fields (DVFs) from the end-of-exhale to other 4DCT phases were computed using deformable image registration. VRT and two ~30s FLs were acquired pre- and post-treatment delivery. We used a simulated-annealing optimization to estimate optimal lung DVFs by maximizing the image mutual information (MI) score between digitally reconstructed radiographs of model-estimated 3D-images and corresponding FLs. The SMM was trained using FL and VRT data from the first breathing-cycle, with the remaining data used for validation. Performance was evaluated using the MI score and Hausdorff distance (HD) of manually contoured anatomical landmarks, computed with respect to FL.
Results: For landmarks near the diaphragm, HD was consistently reduced (mean=5.7mm, maximum=10.6mm) over 4DCT (mean=13mm, maximum=28mm). Using SMM instead of 4DCT improved mean and minimum MI scores by as much as 16% and 10%, respectively, indicating a closer match between model estimation and FLs over 4DCT. Several instances required a model update, without which the performance decreased. The updated model improved mean and minimum MI by up to 19% and 21%.
Conclusion: Integrating additional fluoroscopic images to frequently update the SMM preserved model fidelity. Our SMM consistently out-performed 4DCT for target position estimation, and generated images that more closely matched reference fluoroscopic images.
Funding Support, Disclosures, and Conflict of Interest: This work has been partially supported by the UMBC dissertation fellowship award. Our methodology is demonstrated using patient data collected as a collaborative effort by the faculty and staff at the University of Maryland Medical Center. Data collection was partially supported by NIH-R01-CA 169102.