Training and validation of a knowledge-based dose-volume histogram predictive model in the optimisation of intensity-modulated proton and volumetric modulated arc photon plans for pleural mesothelioma patients

Abstract

Background: To investigate the performance of a narrow-scope knowledge-based RapidPlan (RP) model for optimisation of intensity-modulated proton therapy (IMPT) and volumetric modulated arc therapy (VMAT) plans applied to patients with pleural mesothelioma. Second, estimate the potential benefit of IMPT versus VMAT for this class of patients.

Methods: A cohort of 82 patients was retrospectively selected; 60 were used to "train" a dose-volume histogram predictive model; the remaining 22 provided independent validation. The performance of the RP models was benchmarked, comparing predicted versus achieved mean and near-to-maximum dose for all organs at risk (OARs) in the training set and by quantitative assessment of some dose-volume metrics in the comparison of the validation RP-based data versus the manually optimised training datasets. Treatment plans were designed for a prescription dose of 44 Gy in 22 fractions (proton doses account for a fixed relative biological effectiveness RBE = 1.1).

Results: Training and validation RP-based plans resulted dosimetrically similar for both VMAT and IMPT groups, and the clinical planning aims were met for all structures. The IMPT plans outperformed the VMAT ones for all OARs for the contra-lateral and the mean and low dose regions for the ipsilateral OARs. Concerning the prediction performance of the RP models, the linear regression for the near-to-maximum dose resulted in D_achieved = 1.03D_predicted + 0.58 and D_achieved = 1.02D_predicted + 1.46 for VMAT and IMPT, respectively. For the mean dose it resulted: D_achieved = 0.99D_predicted + 0.34 and D_achieved = 1.05D_predicted + 0.27 respectively. In both cases, the linear correlation between prediction and achievement is granted with an angular coefficient deviating from unity for less than 5%. Concerning the dosimetric comparison between manual plans in the training cohort and RP-based plans in the validation cohort, no clinical differences were observed for the target volumes in both the VMAT and IMPT groups. Similar consistency was observed for the dose-volume metrics analysed for the OAR. This proves the possibility of achieving the same quality of plans with manual procedures (the training set) or with automated RP-based methods (the validation set).

Conclusion: Two models were trained and validated for VMAT and IMPT plans for pleural mesothelioma. The RP model performance resulted satisfactory as measured by the agreement between predicted and achieved (after full optimisation) dose-volume metrics. The IMPT plans outperformed the VMAT plans for all the OARs (with different intensities for contra- or ipsilateral structures). RP-based planning enabled the automation of part of the optimisation and the harmonisation of the dose-volume results between training and validation. The IMPT data showed a systematic significant dosimetric advantage over VMAT. In general, using an RP-based approach can simplify the optimisation workflow in these complex treatment indications without impacting the quality of plans.

Keywords: Intensity-modulated proton therapy; Knowledge-based planning; Machine learning; Pleural mesothelioma; RapidPlan; Volumetric modulated arc therapy.