2081 - The Use of Iterative Knowledge-Based Planning Dosimetry Methods to Improve Head and Neck Cancer Radiation Planning

Purpose/Objective(s): Continuous clinical practice improvement is a standard ABR criteria in Radiation Oncology. Knowledge-based planning (KBP), a form of machine learning relying on pre-existing treatment plans, is a recent advancement in radiotherapy that has improved planning quality, efficiency, and consistency. Institutions can generate KBP models in-house or integrate publicly shared models into clinical workflows to automate planning. Prior to clinical use, it is paramount to assess the performance of KBP models. In our study, we performed a blinded physician review, comparing volumetric modulated arc therapy (VMAT) plans for head and neck (HN) cases that were optimized using manual optimization, in-house KBP, and new publicly available KBP techniques.

Materials/Methods: Ten previously treated patients were retrospectively planned with in-house and Bilateral HN KBP models. Original and KBP optimized plans followed protocol guidelines, delivering 70 Gy, 63 Gy, and 56 Gy to three target levels in 35 fractions. For each patient, four plans were deidentified for physician review: original (Manual), in-house KBP (IH), and Varian KBP with (VarM) and without (VarA) manual assistance during optimization. Four physicians scored all plans from 1 (best) to 4 (worst). Dosimetric scorecards (DSCs), based on multiple piecewise linear scoring functions, were used to assess plans against DVH metrics adopted from NRG HN004.

Results: VarM plans ranked the highest amongst blinded physician review (1^st in 23 (57.5%) and 2^nd in 15 (37.5%)). The second highest ranked was VarA, receiving 7 1^st (17.5%) and 17 2^nd (42.5%) place votes. While KBP can automate planning, integrating manual manipulation during optimization showed VarM improved plan quality to a greater extent than simply running the KBP model automatically. IH and Manual plans yielded just 6 and 4 1^st place votes, respectively; Manual ranked 4^th in 55% of votes (22/40). DSCs corroborated physician rankings, with VarM (82.3±5.0%) outscoring VarA (80.6±5.7%), IH (80.5±5.8%), and Manual (70.1±12.0%), on average. Compared to manual planning, VarM showed significant improvements in terms of dose reduction to larynx (Dmean - 7.1 Gy), spinal cord PRV (D0.03cc – 8.8 Gy), brainstem (D0.03cc – 8.4 Gy), and contralateral parotid-PTV (Dmean - 5.4 Gy); VarM consistently improved PTV coverage to all three prescription levels compared to manual optimization (p<0.05).

Conclusion: Results from blinded physician review and DSC analysis favored the utilization of the Varian KBP model with manual assistance for clinical use in HN VMAT planning. Our clinics have adopted the Varian HN KBP model for clinical use based on the preliminary results of this study.