Model based patient pre-selection for intensity-modulated proton therapy (IMPT) using automated treatment planning and machine learning.

ABSTRACT

BACKGROUND AND PURPOSE: Patient selection for intensity modulated proton therapy (IMPT), using comparative photon therapy planning, is workload-intensive and time-consuming. Pre-selection aims at avoidance of manual IMPT planning for patients that are in the end ineligible. We investigated the use of machine learning together with automated IMPT treatment planning for pre-selection of head and neck cancer patients, and validated the methodology for the Dutch model based selection (MBS) approach. MATERIALS & METHODS: For forty-five head and neck patients with a previous MBS, an IMPT plan was generated with non-clinical, fully-automated planning. Dosimetric differences of these plans with the corresponding previously generated photon plans, and the outcomes of the former MBS, were used to train a Gaussian naïve Bayes classifier for MBS outcome prediction. During training, strong emphasis was placed on avoiding misclassification of IMPT eligible patients (i.e. false negatives). RESULTS: Pre-selection with the classifier resulted in 0 false negatives, 12 (27%) true negatives, 27 (60%) true positives, and only 6 (13%) false positive predictions. Using this pre-selection, the number of formal selection procedures with involved manual IMPT planning that resulted in a negative outcome could be reduced by 67%. CONCLUSION: With pre-selection, using machine learning and automated treatment planning, the percentage of patients with unnecessary manual IMPT planning for MBS could be drastically reduced, thereby saving costs, labor and time. With the developed approach, larger patient populations can be screened, and likely bias in pre-selection of patients can be mitigated by assisting the physician during patient pre-selection.