Benchmarking Publicly Available Knowledge-Based Planning Models with An In-House Trained Model
J Zhang*, J Wolf, K Luca, O Kayode, J Bradley, X Yang, T Liu, J Roper, Emory University, Atlanta, GA
Presentations
TU-D930-IePD-F4-4 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 4
Purpose: Knowledge-based planning (KBP) has demonstrated improvements in clinical treatment plan consistency and efficiency. However, training KBP models is time-consuming and requires numerous high-quality prior plans. While pre-trained public models may offer substantial time savings over in-house models, relative performance is unknown. The purpose of this work is to quantitatively evaluate public and in-house RapidPlan models.
Methods: We evaluated the performance of two publicly available RapidPlan models in comparison with our in-house RapidPlan model and corresponding clinical plans. Thirty previously treated bilateral head-and-neck (HN) patients were randomly selected for validation purposes. None were used for training. RapidPlan models were directly applied to these cases using the same plan settings, including arc geometry, machine parameters, and optimization settings (convergence mode off, MR restart level 3). The new plans were normalized the same as the clinical plans.
Results: The in-house model outperformed the two public models in terms of mean brainstem Dmax (clinical: 18.8Gy, in-house: 18.4Gy, public1: 29.4Gy, public2: 37.4Gy), right parotid Dmean (clinical: 24.4Gy, in-house: 24.5Gy, public1: 27.4Gy, public2: 32.4Gy), left parotid Dmean, (clinical: 24.7Gy, in-house: 23.1Gy, public1: 26.2Gy, public2: 31.9Gy), oral cavity–CTV Dmean (clinical: 41.0Gy, in-house: 38.8Gy, public1: 46.1Gy, public2: 47.4Gy), and spinal cord Dmax (clinical: 20.7Gy, in-house: 26.3Gy, public1: 41.8Gy, pulbic2: 46.5Gy). There are two likely factors for these plan quality differences: (1) planning knowledge progression since the public models were published, and (2) systematic differences in contouring and planning preferences between institutions.
Conclusion: The results demonstrate that our in-house HN KBP model generates treatment plans that are more consistent with our clinical planning standards, as compared with pre-trained public models. While the public models have been previously validated, they may not provide the best plan quality in another patient cohort. Therefore, it is important to validate external KBP models with local datasets prior to clinical implementations.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by Winship Cancer Institute #IRG-21-137-07 -IRG from the American Cancer Society.
Keywords
Treatment Planning, Dose Volume Histograms, Data Interpolation
Taxonomy
TH- External Beam- Photons: Quality Assurance - IMRT/VMAT
Contact Email