Automated workflow for target volume propagation in online adaptive magnetic resonance guided radiotherapy for lung lesions using foundation models.

Abstract

Background and purposeThis study presents a novel workflow for automated gross tumor volume (GTV) adaptation in online adaptive magnetic resonance imaging (MRI)-guided radiotherapy using promptable segmentation foundation models and evaluates its impact on the quality of adapted treatment plans.Materials and methodsTreatment planning GTV contours in lung cancer patients, treated at an MRI-Linac, were propagated to daily MRI using affine or deformable image registration. Subsequently, propagated contours were refined using MedSAM2 or nnInteractive foundation models. The prompting strategy was optimized using an internal validation dataset (16 patients with single GTV, 143 fractions), and its accuracy was evaluated on internal and external datasets (total of 39 patients with single GTV, 225 fractions). In the dose-volume histogram (DVH) evaluation, daily treatment plans were reoptimized to GTVs and planning target volumes (PTV) generated by the proposed workflow. The resulting DVHs were compared with the originals.ResultsNnInteractive improved the segmentation accuracy significantly compared to affine- and TransMorph-based propagation methods. The Dice similarity coefficient improved by up to 15% compared to affine-based propagation, but only marginal improvements were observed compared to TM-based propagation. The DVH evaluation revealed no systematic degradation of plan quality. Most DVH metrics did not differ significantly from the original treatment plans. In 17 out of 21 cases, 100% of the clinically approved GTV volume remained covered by the prescribed dose after the reoptimization.ConclusionThe proposed workflow is feasible and enables automated daily GTV adaptation with improved segmentation accuracy and comparable quality of adapted treatment plans.