Abstract
BACKGROUND: The clinical benefits of recognizing errors from dosimetric quality assurance (DQA) can be realized by improving the dose delivery accuracy. However, an efficient error detection method for data with multiple types of errors is still needed. This study sought to develop an algorithm for quantitatively analyzing multiple errors in DQA data by leveraging Bayesian optimization (BO) and statistical methods. METHODS: The analysis included 79 treatment plans, randomly divided into a training subset (comprising 60 plans) and a testing subset (comprising 19 plans), delivered using an Infinity linear accelerator (LINAC). The analysis examined errors stemming from bilateral multi-leaf collimator (MLC) leaf-banks, jaws, and collimator rotation. A Gaussian process (GP) model functioned as the surrogate for BO, which aimed to adjust the error matrix to minimize failure rates in the DQA. The algorithm's performance was evaluated using simulated and real-world data. To evaluate the efficacy of the algorithm in detecting errors, error matrices of two magnitudes were introduced into the simulations: [-0.5 mm, 0.5 mm, 0.5 mm, 0.5 mm, -0.5 degrees], and [-1 mm, 1 mm, 1 mm, -1 mm, -1 degrees]. In the analysis of the real-world data, inherent systematic errors in the training subset were identified by statistically analyzing the coefficient of variation in the solution sets produced through BO, and corrections were subsequently applied to the original plans. The precision of the error identification was measured by comparing the adjustments to the failure rates for both the training and testing subsets. RESULTS: Systemic biases were identified, and the detected error matrices of [-0.46±0.466 mm, 0.47±0.477 mm, 0.23±1.589 mm, -0.01±1.786 mm, -0.54±0.408 degrees], and [-0.92±0.553 mm, 0.83±0.453 mm, 0.95±1.924 mm, -0.55±1.719 mm, -0.91±0.435 degrees] closely mirrored the expected magnitudes. The analysis of inherent errors revealed substantial improvements in the failure rates following correction, including reductions from 6.06%±4.783% to 1.78%±1.033% in the training subset and from 4.15%±2.643% to 2.02%±1.261% in the testing subset. CONCLUSIONS: The error pattern recognition algorithm can quantitatively detect errors in data with multiple types of errors and analyze the inherent systematic errors in plans that have already passed gamma analysis. The method can enhance the overall performance of plan implementation on specific equipment. Additionally, the algorithm can analyze inherent systematic deviations in clinical DQA data and provide well-labeled datasets for deep-learning methods.