Session Item

Intensity modulated proton therapy (IMPT) is becoming increasingly available worldwide but treatment slots remain a limited resource. An efficient and cost-effective decision support system for patient treatment selection (i.e. photon or proton treatments) is therefore required
The aim of this study was to develop an automated treatment planning method for robustly optimized IMPT plans for oropharyngeal carcinoma patients (OCP) and to compare the results with manual robust IMPT plans.

A random forest based machine learning (ML) model for dose prediction was trained on CT-scans, contours and dose distributions of robust IMPT plans of 81 OCP. The ML model was combied with a robust voxel- and dose volume histogram-based dose mimicking optimizations algorithm for 21 perturbed scenarios to generate a machine deliverable plan from the predicted dose distribution. Next, the ML model and mimicking optimization algorithm was tuned and validated before clinical introduction. Tuning involved adjustments to the mimicking optimization, to generate higher quality machine learning optimization (MLO) plans. MLO tuning was performed using a cross-validation approach with 3x8 tuning patients. Independent testing of the MLO algorithm was performed on another 25 patients. MLO plans were considered clinically acceptable when; the clinical target volume  D98% voxel-wise minimum dose >94% (using 28-perturbed scenario dose evaluations), the conformity index increased < 10 pp (percentage point) and the Normal Tissue Complication Probability (NTCP) (sum of grade-2 dysphagia and xerostomia) increased <2 pp compared to the manual plan. A Bonferroni corrected Wilcoxon signed rank test was performed to assess if the MLO IMPT plans differed statistically significant from the manual IMPT plans with p=0.007 (α=0.05/7 parameters) for target coverage, p=0.006 (α=0.05/9 structures) and p=0.025 (α=0.05/2) for NTCP.  

Adequate robustness was achieved in 24/25 manual plans and 23/25 plans MLO plans in CTV7000. In CTV5425, 22/25 manual plans and 24/25 MLO plans passed the robustness evaluation threshold. The average Σgrade 2 and Σgrade 3 NTCPS were comparable in the manual plans (Σgrade 2 NTCPs: manual. 47.49% vs MLO 48.40%, Σgrade 3 NTCPs: manual 11.89% vs MLAP 12.25%) (figure 1, table 1). The MLO automatically generated deliverable IMPT plans within 45 minutes on average including robustness evaluation.

The combination of machine learning based dose prediction and robust dose mimicking optimization can be used to automatically create clinically acceptable robust IMPT plans for OCP non-inferior to manual treatment plans.