Room: Stars at Night Ballroom 1
Purpose: The aim of the study is to develop a knowledge-based tradeoff hyperplane navigation model to facilitate efficient clinical treatment planning and decision-making in head and neck treatments.
Methods: We propose a model-based approach to predict the tradeoff hyperplane that allows users to navigate the best achievable dosimetric parameters pre-planning. First, using a novel morphological feature, we select a reference pool consisting of similar prior plans from the clinical database. Similar to the conventional knowledge-based planning (KBP) routine, the model is used to estimate the expected dose-volume histogram (DVH) for the current validation case. The discrepancies between the in-model predictions and the actual DVHs are calculated and subsequently processed to extract case-specific tradeoff directions. To evaluate the effectiveness of the proposed tradeoff model, we retrieved 244 anonymized plans and randomly selected 200 plans as the training set, and used the remaining 44 plans as the validation set. We re-planned 14 randomly selected validation cases in various tradeoff locations generated by our proposed model.
Results: The proposed trade-off hyperplane with three directions account for 68.9% Â± 0.5% of the variances in the training plans, and 57.5% Â± 3.0% in the validation plans. Therefore, the predicted best achievable DVH hyperplanes sufficiently explain the discrepancy between the predicted best achievable plans and the clinical plans. All 14 re-planed cases match closely to the predicted hyperplane location (DVH RMSE<10%) and thereby validate the hypothesis that hyperplane-navigated DVHs are achievable for head and neck cases.
Conclusion: The proposed model accurately predicts the best achievable tradeoff hyperplane. It can be incorporated into the KBP framework and produce various plans for clinicians to evaluate. It can also serve the purpose of providing real-time estimations of best achievable plans based on clinical judgment.
Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by a grant from NIH/NCI under grant number R01CA201212 and a master research grant from Varian Medical Systems.