Abstract
This study aimed to evaluate the clinical validity of a dose-mimicking automated planning for volumetric-modulated arc therapy (VMAT) in patients with stage III non-small cell lung cancer (NSCLC), through direct comparison with a commercial dose volume histogram (DVH)-predicted system. We retrospectively analyzed volumetric-modulated arc therapy plans from 75 patients with stage III NSCLC treated at our institution (60 for training, 15 for testing). The dose-mimicking method was implemented using RatoGuide, and the DVH-predicted method was implemented using RapidPlan. The RatoGuide 3D dose-prediction model was trained on the 60 training cases. For each test case, a predicted dose distribution was generated and converted to a deliverable plan (RGDose) in Eclipse using vendor-provided objective functions. A RapidPlan model trained and generated deliverable plans (RPDose) for the same dataset. The clinical plan dose distribution (CliDose) was the reference. We compared dose distributions and DVH parameters among RGDose, RPDose and CliDose. Mean absolute errors (MAEs) relative to CliDose were 0.83 ± 0.66% (targets) and 2.06 ± 3.14% (organs at risk [OARs]) for RGDose, and 0.88 ± 0.66% (targets) and 2.49 ± 3.63% (OARs) for RPDose. There were no significant differences in OAR DVH parameters between RGDose and CliDose. In contrast, compared to CliDose, RPDose showed a significant reduction in the Esophagus D1cc and a significant increase in the Lungs V5Gy. The dose-mimicking method more faithfully reproduced the original clinical plans than the conventional DVH-predicted system, suggesting that dose-mimicking method can capture complex inter-OAR trade-offs and consistently reflect planner intent.