Author: Gregory A Azzam, Michael Butkus, Nesrin Dogan, Robert Kaderka, Nhan Vu, Yihang Xu 👨🔬
Affiliation: University of Miami, Department of Radiation Oncology, University of Miami, University of Miami Sylvester Comprehensive Cancer Center, University of Miami, Sylvester Comprehensive Cancer Center 🌍
Purpose: Knowledge-based planning (KBP) has demonstrated potential for improving planning quality and efficiency. Adoption of KBP in particle therapy has been slow due to limited number of plans to train models. This study develops a brain KBP routine for intensity-modulated proton therapy (IMPT) and investigates common questions when there is limited training data: Can the training set be used for validation and is there a major improvement with larger models?
Methods: 120 IMPT plans for brain cancer were identified, with 31 selected for validation. Three KBP models were trained: "KBPValidation" (31 plans), "KBPOpen-loop" (89 plans excluding KBPValidation), and "KBPClosed-loop" (all 120 plans) with identical optimization objectives. The 31 validation plans were re-planned using each KBP model. Plans were compared dosimetrically to Human-generated clinical plans and each other using DVH parameters and paired t-tests(p<0.05). Additionally, blinded physician-review compared Human and KBPClosed-loop plans.
Results: Minimal differences were observed in plans generated by Humans and the three KBP models. All KBP models significantly reduced CTV V105% (Average: 15.2%/3.7%/4.3%/4.5% for Human/KBPValidation/KBPOpen-loop/KBPClosed-loop). Compared to Human Plans, KBPValidation reduced Hippocampus_L Dmean (535cGy/790cGy) and both KBPOpen-loop&KBPClosed-loop reduced pituitary Dmax (1992cGy/1761cGy/1801cGy). Chiasm Dmax was reduced in KBPOpen-loop versus KBPClosed-loop (2813cGy/2912cGy). No other DVH parameters were significantly different.
In the blinded physician-review, all KBPClosed-loop plans were considered clinically acceptable. KBPClosed-loop was preferred in 12/31 cases, Human plans 6/31 times and a tie for 13/31 plans.
Conclusion: The KBP models generated plans with dosimetric quality comparable to Human-planned treatments. Minimal differences between KBP model sizes suggests that clinics can train and validate on the same plans if there is only a sparse amount of plans available. High acceptance rate of the KBP plans in blinded physician-review confirms the performance of the KBP models. These findings facilitate the implementation of KBP in proton therapy where often only limited data is available.