Automated Machine-Learning Radiation Therapy Treatment Planning for Pediatric and Adult Brain Tumors

Abstract

To evaluate the feasibility of using a machine-learning (ML) treatment planning method to automatically generate treatment plans for pediatric and adult brain tumor cases, and compare ML-generated treatment plans with human-generated plans used clinically. The application of a ML radiation treatment planning method [1] was investigated using a training set of adults (n = 25) and pediatric (n = 91) brain tumor patients treated from December 2007 to March 2018 with dose prescription of 54 Gy in 30 fractions using IMRT/VMAT. Our ML model predicts the optimal dose to targets and normal tissues based on learned relationships between image features and dose distributions of anatomically similar patients. Predicted dose plans are converted to clinically deliverable single-arc VMAT plans using optimization algorithms that enforce technical beam delivery constraints. Ten children and three adults with centrally-located primary brain tumors clinically treated with 54 Gy, not previously used for model training, were re-planned with this ML model. Dosimetry to targets and organs-at-risk (OAR) were compared between ML and manual clinical plans. Comparable target coverage was observed in both ML and manual plans. At least 95% of PTV received >51.3 Gy (95% of prescription) in 11 ML vs 12 manual plans. Similar maximum doses to OARs including optic nerves, chiasm and cord were observed. Maximum dose to optic nerves was <54 Gy in all ML and manual plans; maximum chiasm dose was <54 Gy in 12 ML vs 13 manual plans; maximum cord dose was <54 Gy in 12 ML vs 11 manual plans. Maximum brainstem dose was >54 Gy in 11 out of 13 ML (85%) and 7 manual (54%) plans, because 81% of the training cases with centrally located tumors had brainstem doses >54 Gy. The median differences in mean dose to left and right cochleae were 3 Gy and 0.3 Gy higher in ML than manual plans, respectively, whereas mean doses to left and right temporal lobes were 2.5 Gy and 2.8 Gy lower in ML plans than manual plans, respectively. Average time for ML plan generation was 27 minutes (range, 17-34). We demonstrate the feasibility of rapidly generating clinically-deliverable ML plans with consistent plan quality, similar target coverage and OAR sparing as compared to manual clinical plans. A similar proportion of cases with maximum brainstem doses >54 Gy was observed in both ML plans and historical training cases which illustrates the importance of continuous quality improvement in ML models.