2046 - Feasibility of DLAS-Driven Reoptimization to Accelerate MR-Guided Online Adaptive Radiotherapy for Pancreas Malignancies

Purpose/Objective(s): Treatment times for MR-guided online adaptive radiation therapy (MRgOART) remain suboptimal, particularly for abdominal malignancies. Currently, physicians are asked to manually review contours on daily images and then manually review daily adaptive plans. This two-step review process increases physician burden, precluding physicians from attending to other clinical tasks or increasing MRgOART treatment times if physicians leave the procedure between contour and plan review. Deep learning auto-segmentation (DLAS) is quickly emerging as a powerful solution for daily contouring. We investigate here the dosimetric impact of DLAS driven MRgOART, in which daily adaptive plans are generated directly using unedited contours, to determine the feasibility of combining contour review and adaptive plan review to reduce physician burden and accelerate MRgOART for pancreas malignancies.

Materials/Methods: Two pancreas SBRT patients treated using the Adapt To Shape (ATS) workflow on a 1.5T MR Linac were included in this retrospective study. For each of five treatment fractions, target contours were rigidly transferred and stomach, duodenum, small bowel, colon, L kidney, R kidney, and liver were derived using DLAS trained on abdominal MR-Linac images. DLAS driven daily adaptive plans were generated using full fluence reoptimization with no contour modifications and then recalculated on physician-approved contours to compare against clinical ATS plans. Segmentation accuracy was evaluated against ground truth contours using Dice similarity coefficient (DSC) and mean distance to agreement (MDA).

Results: Segmentation accuracy and dosimetric results are shown in the table. No significant differences in GTV or PTV 25 coverage were observed between plans for either patient. No significant difference in PTV 35 coverage was detected for Patient 1. PTV 35 coverage was reduced in the DLAS driven plan for Patient 2 (p = 0.02), however coverage remained within acceptable tolerance. No significant differences were observed in D0.03cc for the luminal GI OARs between the two plans.

Conclusion: DLAS driven MRgOART is becoming feasible. With this acceleration approach, physician burden is reduced, and treatment times can be reduced by combining contour review with adaptive plan review in a single step.