2974 - Streamlining Abdominal MR-Guided Online Adaptive Radiation Therapy with DLAS-Driven Reoptimization

Purpose/Objective(s): Inter-fractional changes in abdominal anatomy can be addressed with MR-guided online adaptive radiation therapy (MRgOART). The adaptive plan requires contour editing of organs at risk (OARs), increasing workflow time. The longer the contour editing, the higher likelihood of intra-fractional motion, which can challenge the dosimetric benefits of adaptive planning. Robust deep learning auto-segmentation (DLAS) of MR images is one way to shorten the time between daily imaging and treatment. This study investigates the potential time savings of DLAS-driven MRgOART by a streamlined process that omits contour editing and combines contour review and plan approval into a single step.

Materials/Methods: Six patients with five fractions each treated for abdominal malignancies on a 1.5T MR-Linac were chosen for this retrospective study. The DLAS-driven plans were created by re-optimizing with unedited DLAS contours generated on the daily OAR anatomy, and targets rigidly transferred from reference images. The dosimetry of the DLAS-driven plan was compared with the clinical plan. Geometric accuracy of DLAS contours versus clinical contours was assessed using the Dice Similarity Coefficient (DSC) and mean distance-to-agreement (MDA). Plan quality was evaluated by target coverage and D0.03cc to the OARs. Wilcoxon matched-pairs signed-rank tests were used to examine dosimetric differences between DLAS-driven and clinical plans. Time savings were estimated based on clinical times for contour editing, plan review, and plan approval.

Results: Most OARs had DSC scores above 0.8 and a mean MDA below 3 mm, indicating strong agreement between DLAS and clinical contours. No notable differences were seen in target coverage or D0.03cc between the clinically delivered plans and those generated by DLAS. Estimated time savings from the DLAS-driven approach ranged from 8 - 24 minutes per fraction, based on patient median clinical contouring times, with the 30 fractions ranging from 6.0 to 38.2 minutes. Clinical times, average DSC for duodenum, and median PTV coverage differences are in Table 1.

Conclusion: In this study, plans generated through DLAS-driven methods were similar to those delivered clinically. Accurate MR-based DLAS models could enable DLAS-driven adaptive planning, reducing the need for contour edits before reoptimization. This method could reduce physician workload and shorten MRgOART treatment times by combining contour review and plan approval into a single step.