Room: Exhibit Hall | Forum 6
Purpose: The goal of radiation therapy (RT) peer review rounds is to identify errors, reduce treatment variation, and ensure plan quality and safety. The ability to assess all cases in a peer review setting is impeded by high workloads and increasing plan complexity. The purpose of this study was to evaluate the ability of a machine learning (ML)-based method to triage RT plans of high complexity and/or with errors for peer review prioritization.
Methods: We assigned binary planning complexity scores to 202 breast RT plans during weekly peer review rounds over 15 consecutive weeks from August to November 2018. Assigned scores were based on whether the treatment plan elicited discussion (1, outlier, n=38) or not (0, normal, n=164) during rounds. An isolation forest algorithm was trained on over 3,000 previously clinically-approved breast RT plans using engineered and learned features of the CT image, delineated regions of interest, dose distribution, and the plan including beams. The 202 scored RT plans were used as an independent testing set to evaluate the ability to detect complex plans that prompted discussion during peer review.
Results: All RT plans were assigned a complexity score between 0 and 1 by the ML algorithm, where higher scores indicated plans with greater deviations from the training set. The area under the curve (AUC) was 0.63 with a 95% confidence interval of [0.51, 0.74]. Misclassified plans were investigated to determine the strengths and pitfalls of the method for future feature and data tuning.
Conclusion: This machine-learning based method has the potential to improve the efficiency of peer review by augmenting human discussion through case prioritization. Ongoing work to improve performance of this framework includes the addition of clinical features such as margins and grade, and the development of interpretable non-binary scores to reflect degree of complexity for case ranking.