Externally validated knowledge-based planning model for machine learning-assisted radiation therapy treatment of
Samantha Am Lloyd1, Carrie-Lynne Swift2, Andrew G Pritchard3
1BC Cancer - Vancouver, Medical Physics, Vancouver, British Columbia, V5Z 4E6; University of British Columbia, Division of Radiation Oncology, Vancouver, British Columbia, V5Z 1M9.
None:
We have trained and externally validated a knowledge-based planning model for radiation therapy planning in the setting of high-grade glioma. Model performance and utility in the context of clinical trial radiotherapy quality assurance (RTQA) are presented. A RapidPlan (RP) model was trained on 65 cases and tested on an additional 20 cases that were manually optimized and delivered within our institution. The model was externally validated on 34 cases that were manually optimized and treated at four outside institutions. These cases were selected to have target overlap with the brainstem or optic pathways. RapidPlan-generated plans were evaluated against planning objectives and manually optimized clinical plans. Cases were classified as (1) Clinical plan was superior, (2) Clinical plan was within 5 % of RP, or (3) Clinical plan could be improved. The clinical plan was characterized as superior for an objective if the metric was >5 % better than in the RP plan or if the clinical plan met the objective but the RP plan did not. Possible clinical plan improvement was indicated for an objective if the metric was >5 % better in the RP plan, or if the RP plan met the objective but the clinical plan did not. A Wilcoxon signed-rank test with a p < 0.05 significance threshold was used to determine if differences in PTV coverage, OAR doses and MU were statistically significant. Eight of 34 RP plans met all planning objectives in a single optimization, while an additional six met all normal tissue objectives while compromising target coverage. In more than 80 % of cases, when an objective was not achieved by RP, it was also not achieved in the manually optimized plan. Comparisons of clinical plans and RapidPlan indicated that statistically significant improvement was possible for Optic Nerve Dmax and Optics PRV V54; however, RP also introduced a statistically significant increase in Dmax overall. Clinical plans could have been improved for individual planning objectives 24 to 68 % of the time, while the clinical plans were considered superior to RP for individual planning objectives 9 to 51 % of the time. Improvement in PTV coverage was possible for 18 % of clinical plans. The presented RapidPlan model for high-grade glioma performs well for cases both within and outside our institution. The model has demonstrated the capacity to reduce normal tissue dose to optic structures and to provide feedback in the context of clinical trial plan RTQA.
