Feasibility of DLAS-Driven Reoptimization to Accelerate MR-guided Online Adaptive Radiotherapy for Abdominal Malignancies.
Authors
Affiliations (4)
Affiliations (4)
- Medical College of Wisconsin, Milwaukee, WI, USA. Electronic address: [email protected].
- Medical College of Wisconsin, Milwaukee, WI, USA.
- Elekta AB, Stockholm, Sweden.
- Medical College of Wisconsin, Milwaukee, WI, USA. Electronic address: [email protected].
Abstract
MR-guided online adaptive radiation therapy (MRgOART) adjusts for daily anatomical changes. One challenge is prolonged treatment time due to organs-at-risk (OARs) contour editing. Recently, deep learning auto-segmentation (DLAS) models have demonstrated accuracy on MRI images, reducing contour editing demands. In this retrospective study, we investigate the dosimetric impact of DLAS-driven MRgOART, in which daily adaptive plans are created directly using unedited DLAS contours, and the feasibility of a one-step contour and adaptive plan review and approval process. Six patients treated for abdominal malignancies on a 1.5T MR-Linac were studied. Daily OAR anatomy was segmented by a MR-based DLAS model, with targets rigidly transferred from reference images. The DLAS-driven plan was generated by re-optimizing using unedited DLAS contours. Dose was recalculated using clinical contours for dosimetric comparison with the clinical plan. Dice Similarity Coefficient (DSC), mean distance-to-agreement (MDA), surface Dice Similarity Coefficient (sDSC), and relative added path length (rAPL) were used to compare geometric accuracy of DLAS to clinical contours. Target coverage, D0.03cc to the gastrointestinal OARs, and V1200cGy to kidneys and liver were used to assess plan quality. Wilcoxon matched-pairs signed-rank tests evaluated dosimetric differences between DLAS-driven and clinically delivered plans. Time savings for contour editing to plan approval were estimated. DLAS and clinical contours demonstrated good agreement, with most average DSC scores exceeding 0.8 and average MDA under 3mm for OARs. No significant differences were found in target coverage, D0.03cc, or V1200cGy between the clinically delivered and DLAS-driven plans. Estimated treatment time savings ranged from 8 to 24 minutes. Accurate MR-based DLAS models may enable DLAS-driven adaptive planning, eliminating contour edits prior to reoptimization. DLAS-driven plans were comparable to clinically delivered plans for the six patients studied. DLAS-driven adaptive planning is a promising approach to reduce physician burden and treatment times in MRgOART.