2022 - Evaluation of Iterative Deformable 2D-3D Registration for Motion Management with Pancreatic SBRT Patients

Purpose/Objective(s): Accurate motion management is vital in pancreatic stereotactic body radiotherapy (SBRT) to ensure ablative doses are confined to target volumes as pancreatic targets are subject to complex respiratory, peristaltic and digestive motions. For standard linacs with onboard imagers, deformable 2D-3D registration could be used to register planning CTs or setup CBCTs to intrafraction kV images and monitor intrafractional anatomy. In this work, we assess the hypothesis that deformable 2D-3D registration can monitor simulated pancreatic and duodenal motions with clinically acceptable accuracy in =50% of test cases.

Materials/Methods: Planning data (4DCT and a breath hold or gated 3DCT) from 15 patients was collected during an ethics approved pancreas SBRT trial. Deformable registration was used to register each patient’s planning CT to each phase of their 4DCT (referred to as ground truth registrations). For each patient, their planning CT and structure set were deformed by their ground truth registrations. Digitally reconstructed radiographs (DRRs) were generated from each deformed CT at (0, 90, 180 and 270)° projection angles, with 600 DRRs generated in total. The deformable 2D-3D registration algorithm was applied to register each patient’s planning CT to their corresponding DRRs. In the algorithm, 3D anatomical motion was modelled using a 3D cubic BSpline transform and gradient ascent maximized the Pearson’s Correlation between the input DRR and the DRR generated from the BSpline transformed CT. Registration accuracy was assessed by comparing planning contours deformed by the 2D-3D registration to those from ground truth registration. Surface dice similarity coefficient (DSC; using 3 mm tolerance) and mean distance to agreement (MDA) were computed for the gross tumor volume (GTV), pancreas head, pancreas and duodenum. For each structure, registrations were deemed clinically acceptable when DSC = 0.7 and MDA = 2 mm.

Results: Each structure’s median and [5%, 95%] for DSC and MDA, and acceptance rate are reported in Table 1. For all structures, acceptance rates were between 70% and 77%. The mean registration time was 7 minutes, 41 seconds.

Conclusion: The proposed approach monitored the simulated pancreatic and duodenal motions with clinically acceptable accuracy in =50% of test cases. Deformable 2D-3D registration may enable 3D motion management from triggered 2D images with standard linacs.