Reducing IMRT QA Workload by 95% and Keeping the Same Level of Quality Control

Abstract

Reducing the maximum possible error (max-error) in a quality assurance program is the natural metric in contrast to reducing the mean value of the expected error. Current machine learning (ML) models used to predict IMRT QA, however, minimize the later and not the former. Here, we propose a novel ML algorithm for predicting QA plan passing rate that minimizes the max-error. From a single institution, 498 patient IMRT QA plans were measured using a commercial 2D diode array (Sun Nuclear’s MapCheck®) and delivered on five isocentric linear accelerators. Passing rates were determined based on 3%/3mm local dose and distance-to-agreement. Each treatment plan was characterized by 78 features to quantify complexity, such as MLC position, field area, X/Y jaw position, etc. Two ML models were used to predict passing rate: (1) an ordinary least squares (OLS) linear model that minimize the mean of the expected error; and (2) Chebyshev minimax (MM) linear solution that minimizes max-error. An 80% training dataset was used for fitting and optimization, and a 20% testing dataset was used to compare predictions. Model fitting and optimization (using Ridge, Lasso and Lasso-like regularization) was performed in a computer algorithm. When using all 498 plans for training, the max-error of IMRT QA passing rate predictions was 7.6% for the OLS model and 3.0% for the MM model. The mean square error was, however, 1.4% and 2.8% for the OLS and MM models respectively showing the tradeoff between both optimization methods. Following optimization and testing with a dataset that the model had not seen before (20% plans), the OLS and MM model max-errors were 7.0% and 3.8% respectively. To reach the clinical goal that all plans have at least a 90% passing rate, using the OLS we would have to QA all plans predicted to be 97.0% passing rate or lower (79 out of 497, 16.0%). Whereas with the MM model, we would only need to QA all plans predicted to be 93.8% or lower (24 out of 497, 4.8%) which is a 4x reduction compared to the OLS model. Efficient and safe QA programs for IMRT treatments require accurate identification of plans that are likely to not meet passing criteria. Such a program would enable relocation of clinical resources to higher priority tasks, which is in alignment with TG100 guidelines. Using Chebyshev minimax optimization, we could reduce resources for IMRT QA by 95% while guaranteeing a safe clinical QA program that meets current passing criteria.