Investigation of Knowledge-Based Planning Models for Intensity-Modulated Proton Therapy With Limited Data for Brain Targets

Background Knowledge-based planning (KBP) in radiation therapy aims to improve the plan quality and planning efficiency. Limited training data complicate the development of KBP models for intensity-modulated proton therapy (IMPT). Clinicians face questions such as whether KBP validation is feasible on the training set and if small data sets substantially impact the resulting plan quality. This work describes the development of KBP models for brain IMPT while focusing on these key questions. Materials and methods A hundred and twenty brain IMPT plans were identified and 31 were selected for validation. Three KBP models were trained (a) KBP_V trained on the 31 validation plans; (b) KBP_O, an open-loop model with 89 plans excluding the validation set; and (c) KBP_C, a closed-loop model combining all the 120 plans. The 31 validation cases were re-planned using each KBP model. The resultant KBP plans were compared to manually-generated clinical plans and each other using dosimetric parameters and paired t-tests (p<0.05). Additionally, the manual and KBP_C plans were evaluated through a blind physician review. Results Dosimetric differences between manual plans and the three KBP models were minimal. The KBP_V, KBP_O, and KBP_C models significantly reduced the clinical volume receiving 105% of the prescribed dose (V105%) to an average of 3.7%, 4.3%, and 4.5%, respectively compared to manual plans showing 15.2%. KBP_V reduced the mean dose to the left hippocampus (Hippocampus_L D_mean) to 535±630 cGy compared to manual plans showing an average of 790±1077 cGy. The maximum dose to the pituitary gland (pituitary D_max) for the manual plans was 1992±1951 cGy, which was reduced in both KBP_O and KBP_C to 1761±1832 cGy and 1801±1831 cGy, respectively. When the models were compared, KBP_O reduced the maximum dose to the optic chiasm (chiasm D_max) to a greater extent than KBP_C (2813 cGy vs. 2912 cGy, respectively). In the blind physician review, all the KBP_C plans were considered clinically acceptable. When stating a preference, the physician favored the KBP_C plans in 12 out of 31 cases. A tie was scored for 13 out of the 31 plans, while the manual plans were preferred in six out of the 31 cases. Conclusion All KBP models produced plans that were dosimetrically comparable to manually-generated reference plans. The blind physician review confirmed that the plans were clinically viable. Minimal dosimetric differences between the three different KBP models for IMPT suggest that KBP training and validation sets can be identical if needed. This finding supports the generation of KBP models for situations where the training data may be limited.