Applying a pretrained DL-based IMPT planning workflow in an external center: A feasibility study in oropharyngeal cancer.

PURPOSE: This study explores the feasibility of applying DL-based robustly optimized intensity modulated proton therapy (IMPT) planning in patients with oropharyngeal carcinoma (OPC) at an external center. MATERIALS AND METHODS: A U-Net deep learning model was trained to predict dose distributions for robust IMPT plans in RayStation. A mimicking algorithm translated dose predictions into clinically deliverable, deep learning optimization (DLO) plans. Mimicking parameters were reconfigured using five OPC patients towards clinical goals of the external center within two working days. Ten independent cases were used for evaluation of the DLO plans; using a blinded multidisciplinary review and quantitative analysis, including assessments of target coverage, organ at risk (OAR) doses and normal tissue complication probabilities (NTCPs). RESULTS: Manual and DLO plans were each preferred in 40% of cases, while 20% of cases were rated as equivalent. Sixty percent (6/10) of both manual and DLO plans were deemed clinically acceptable. DLO plans significantly reduced the mean dose to the oral cavity (Manual: 31.5 Gy (RBE) vs. DLO: 30.2 Gy (RBE), p = 0.002), and to the parotid glands (Manual: 20.1 Gy (RBE) vs. DLO: 17.8 Gy (RBE), p < 0.001). They also reduced the NTCP for xerostomia grade ≥ 2 (Manual: 38.1% vs. DLO: 37.1%, p = 0.002) and grade ≥ 3 (Manual: 10.3% vs. DLO: 9.9%, p = 0.002) . CONCLUSION: This study demonstrated that deep learning-based IMPT planning for OPC patients generated fully automated IMPT plans with quality comparable to manual planning at an external center.