2358 - Retreatment Review for Abdominal Cancer: Assessing the Impact of Deformable Image Registration and AI Auto-Segmentation on Dose Accumulation

Y. Liu¹, A. Dinh², C. Gui³, S. Elguindi⁴, W. Lu⁵, J. G. Mechalakos⁴, Y. C. Hu⁵, L. Kuo⁴, L. I. Cervino⁴, and P. Zhang⁴; ¹Memorial Sloan Kettering Cancer Center, MIDDLETOWN, NJ, ²The City University of New York, New York, NY, ³Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, ⁴Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, ⁵Memorial Sloan Kettering Cancer Center, New York, NY

Purpose/Objective(s): To assess the use of Deformable Image Registration (DIR) in dose accumulation for abdominal cancer retreatment, employing contours outlined by both physicians and AI auto-segmentation. Materials/

Methods: 8 abdominal cancer patients undergoing retreatment were involved in the study. DIR was implemented from previous to current computed tomography (CT) scans, followed by dose accumulation calculation. The dose accumulation was implemented by an in-house developed script - the Radiotherapy Dose Accumulation Routine (RADAR). As a comparison, dose summation with rigid image registration in the treatment planning system (TPS) were also performed. Within the dose accumulation workflow, a primary challenge was assessing the reliability of DIR. To accomplish this, the physician (MD) drawn contours on the previous CT, deformed by DIR, were compared with the reference contours on the current CT. The reference contours were either delineated by MDs or generated through an in-house developed AI auto-segmentation tool, with our study employing both (MD and AI) seperately. AI-based segmentation was included in the workflow to reduce the clinicians burden and enable a fully automated pipeline for dose accumulation. To pave the way for future research and extend the DIR evaluation to a larger cohort, we also compared our AI contours with those from MDs, focusing on one organs at risk – stomach – for abdomen. 2 DIR algorithms (intensity-based and mutral information based) were evaluated.
Results: The study found that RADAR with DIR provided different dose accumulation for all 8 patients compared with rigid registration by TPS. Dosimetric discrepancies between DIR and rigid registration were significant for Dmax of cord and bowel, and Dmean of liver. Variations were observed but not statistically significant for Dmax of stomach and esophagus. When comparing two DIR algorithms, minimal differences were noted, suggesting consistent DIR performance for dose accumulation. For stomach, MD contours showed an average Mean Distance to Agreement (MDA) on contour surfaces of 6.7±3.0 and 7.9±4.6 mm using the 2 DIR algorithms, respectively, with Dice similarity coefficients of 0.66±0.13 and 0.65±0.18. Similarly, AI contours show an average MDA of 7.3±3.2 and 6.7±4.5 mm using the same 2 DIR, with Dice coefficients of 0.67±0.14 and 0.72±0.27. The differences between AI and MD contours were minimal, suggesting comparable accuracy and potential use of AI in future studies for a larger cohort or clinical dose accumulation workflows.
Conclusion: In this study, the performance of DIR was evaluated for its role in calculating dose accumulation during abdominal cancer retreatment. By quantifying associated uncertainties, DIR may offer enhanced accuracy in dose accumulation compared to conventional rigid registration methods. The comparable DIR accuracy between AI-generated and physician-delineated contours suggests a valuable role for AI in future clinical dose accumulation and planning.