Workflow to detect exceeded dose constraints in pancreatic stereotactic body irradiation after intrafraction motion during magnetic resonance-guided adaptive radiotherapy using a deep learning-refined contour propagation tool.

Introduction: The adapt-to-shape (ATS) process on the MR-Linac involves manual contour edits followed by treatment plan re-optimization on daily pre-beam MRIs. A verification image is acquired after plan optimization to assess the dose distribution with respect to intrafraction motion using the pre-beam contours. We introduce here a workflow to automatically detect organ motion of gastrointestinal structures that results in exceeded planned dose constraints. Materials and methods: The workflow first transferred the contours and dose distribution created on the daily pre-beam MRI to the verification MRI. A deep learning-refined contour propagation (DL-RCP) tool, trained on 79 images, improved the transferred contours and the dose to 0.03 cm3 (D0.03 cm3 ) is updated. The workflow notified the physician if D0.03 cm3 exceeds the constraint. We tested the workflow on 48 daily ATS fractions of 11 patients treated for pancreatic cancer (33-40 Gy in 5 fractions). We added manually drawn contours to the verification images for reference. Results: The Dice similarity coefficient and mean distance to agreement for the clinical/DL-refined contours were 0.56/0.71 and 9.08/5.08 mm, respectively. The workflow detected exceeded constraints with specificity 0.90, sensitivity 0.75, and accuracy 0.85. In one case, the duodenal D0.03 cm3 was 29.9 Gy for the clinical contour, and 36.0 Gy and 35.6 Gy with the reference and DL-refined contours. Conclusion: The proposed method detected exceeded dose constraints in gastrointestinal structures due to intrafraction motion during ATS planning for pancreatic treatments and can aid in the clinical decision to re-optimize the plan on the verification MR.