Go to AFOMP Pulse Contents/Index Page
Min Ma1, Ke Zhang1, Pan Ma1, Zhihui Hu1, Hui Yan1, Kuo Men1, Jianrong Dai1#
1 National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China
# Corresponding authors, Jianrong Dai, Email: dai_jianrong@cicams.ac.cn
The medical electron linear accelerator (Linac) is the widely used external radiotherapy equipment and can meet the needs of most patients. The Linac is the delivery equipment for radiotherapy, delivering radiation in a specific form to the tumor site while avoiding damage to the surrounding normal tissue. Its performance and correct operation have a direct impact on the outcome of tumor treatment. Quality control (QC) is an important means to ensure that the quality status of the Linac is essentially the same as the acceptance and commissioning; by defining the QC standards for the Linac and developing the QC methods; based on the actual situation, exploring the differences between actual and standard, and analyzing the causes; taking action where tolerances or action levels are exceeded. Studies have shown that the failure of Linac is an important source of radiotherapy errors and accidents; timely detection and correction of Linac failures during the deliver phase, prevents causing more patient errors during treatment and improves the safety and efficacy of patient treatment. As can be seen, QC of Linac is an indispensable step in radiotherapy.
The conventional method is to carry out periodic QC testing of Linacs based on QC guidelines/reports issued by national/international associations. This method aims to assess the status of the performance parameters and to detect failures through QC results. However, the QC items and frequency recommended by the conventional method are fixed and do not effectively reflect the status of a Linac; QC is ineffective, and most tests do not detect problems in a timely manner; the work time and workload required by physicists to complete QC is long. However, errors still occur, and various types of accidents are difficult to eliminate.
In recent years, the radiotherapy community has actively introduced new methods of QC from other industries. They hope that analyzing existing QC data is to gain a more accurate understanding of the operational status of radiotherapy equipment, determine the frequency and design tolerances of QC. The new methods based on the time of application on QC are divided into Statistical Process Control (SPC), Failure Mode and Effects Analysis (FMEA), Risk Matrix, Artificial Intelligence (AI), and Six-Sigma Methodology (SSM). The SPC is applied to the QC of Linac to monitor changes in QC data, identify the failures in a timely manner and design tolerance ranges for QC items; FMEA, as one of the currently popular quality management tools, determines the priority and frequency of QC items based on the Risk Priority Number (RPN); Risk matrix is used to determine the level of risk according to a two-dimensional matrix; AI continuously learn data characteristics and predict trends of QC data in Linac; SSM focuses on further process optimization in response to shortcomings in the process. However, there are still some shortcomings in these new methods: all of them are analyzed for a single quality control item, are focused on process application, and have less application for quality control of radiotherapy equipment; besides, the FMEA may be influenced by the subjective experience of experts when determining the scoring of failure modes, thus affecting the objectivity of the QC content; the AI is applied to the QC of radiotherapy equipment, the predictive effect still needs to be improved and it has not really been implemented into clinical applications. In addition, the radiotherapy community should actively introduce new QC methods from industries that do well in quality assurance, such as blood transfusion, anesthesia and clinical chemistry, aviation, and nuclear power plants.
This study proposes to introduce a patient risk model from the clinical chemistry and to improve the new methods already introduced in the radiotherapy. It aims to make QC methods of Linac more objective, quantitative, and more effective, and detecting failures in a timely manner. The main research covers the following four areas:
(1) Patient risk model to determine the QC frequency of Linac
This part of the study is the first to apply a patient risk model to radiotherapy. It aims to determine the QC frequency of the Tomotherapy system, to ensure that no patient outcomes are in error during the operation of the radiotherapy equipment. The patient risk model was divided into three main steps: (i) the power function graph was generated by program simulation to select the optimal QC rule and the number of times (n) each QC rule was evaluated. (ii) The new QC frequency was the smallest integer value of the number of patients treated between QC tests Nb(s). (iii) Prospectively collected QC test data and evaluated for new and traditional QC frequencies using individual control charts (I-Charts). Based on the power function graph, the 13s control rule and n = 5 was selected. Nb(s) decreased and then increased with increasing the systematic error (s). The smallest integer value of Nb(s) was 21, which was the new frequency of output constancy in the Tomotherapy system. In the I-Chart of the new frequency, the out-control point appears at the 29th. In the I-Chart for the conventional frequency, the out-control point appears at the 25th and 37th. Retrospective analysis of the records of failures of the Tomotherapy system during the evaluation period revealed that the new frequency found out-of-control appeared before the failure, while the conventional frequency found out-of-control appeared after the failure. The new frequencies could prioritize the detection of radiotherapy equipment failures over the conventional frequencies. The new frequency is not for individual patient outcomes, but for the average patient outcome treated on a Tomotherapy system.
(2) Risk matrix to determine the QC frequency of Linac
This part of the study applies the risk matrix for the first time to analyze the risk level of QC items and to quantify the frequency of QC. At the corresponding frequency, each QC item exceeding the tolerance corresponded to a failure mode. The failure modes contained three parameters: S, O and D. The S was determined by the impact corresponding to the percentage dose difference between the original plan and the error plan. O was calculated based on the frequency with which QC data exceeded the tolerance. D was the probability that QC data exceeded the tolerance but was not detected. The risk matrix is to apply a two-dimensional matrix of S and O values to visualize the risk areas of the failure modes. It is classified as low, medium, and high risk. In this study, the time corresponding to the first occurrence of medium risk was used as the new QC frequency. The E=O/D metric assessed the performance of the QC frequency. QC data were collected on three conventional Linacs: LN1 (Elekta VersaHD), LN2 (Varian Novalis) and LN3 (Varian Edge). They included 1 dosimetry parameter and 11 mechanical parameters: X-ray output constancy (QC1), Distance indicator @ iso (QC2), Laser localization (QC3), Treatment couch position (QC4), Gantry rotation isocenter (QC5), Couch rotation isocenter (QC6), Collimator angle indicators (QC7), Gantry angle indicators (QC8), Treatment couch position (QC9), Light field coincidence (QC10), MV/kV: imaging and treatment isocenter coincidence (QC11), and Leaf position accuracy (QC12). For LN1, the frequency of QC1 and QC3 was daily; QC2 and QC12 was weekly; QC8 and QC9 was biweekly; QC7 was monthly; QC11 was bi-monthly; and QC4, QC5 QC6 and QC10 had a frequency of annually. For LN2, the frequency was weekly for QC1, QC2, QC3, and QC12; biweekly for QC4; bimonthly for QC9 and QC11; and annual for QC5, QC6, QC7, QC8, and QC10. For LN3, the frequencies of QC1 and QC3 were daily; QC12 was weekly; QC7 and QC8 was bi-monthly; and QC2, QC4, QC5, QC6, QC9, QC10, and QC11 was annual. The E obtained at the new frequency are not lower than those obtained at the conventional frequency, indicating that QC testing at the new frequency can detect equipment failures in advance. The risk matrix was applied to the QC of the three conventional Linacs to quantitatively determine the frequency of QC and provide an effective strategy for the risk level of QC items on radiotherapy equipment.
(3) Six-sigma methodology to design QC limits of Linac
This part of the study introduces that the six-sigma methodology (SSM) personalized design QC limits (tolerance limits and action limits). A framework is highlighted to clarify the various stage. In the define stage, the limits of the QC items need to be defined. In the measure stage, daily QC data were collected retrospectively in the Machine Performance Check (MPC) system. In the analysis stage, statistical analysis and process capability index presented the rationale for how to determine the limit values. In the improve stage, action limits were calculated using the process capability index; tolerance limits were determined using the larger control limits in the individual control chart. In the control stage, daily QC data was prospectively collected; the effect of action limits and tolerance limits were monitored using the I-Charts. Collimation Rotation offset had minimum the process capability index, that is the minimum Cp, minimum Cpk, minimum Pp and minimum Ppk values for 2.53, 1.99, 1.59 and 1.25, respectively. CouchRtn had maximum the process capability index, that is the maximum Cp, maximum Cpk, maximum Pp and maximum Ppk values for 31.5, 29.9, 23.4 and 22.2, respectively. The action limits for the three QCs were higher than the recommended tolerance values, i.e., ISO Center Size, MLC Max OffsetA, and Rotation Induced Couch Shift. The new tolerance limits for all QC items were lower than the original. Some of the data on the I-Charts for Beam Output Change, ISO KV, and JawX1 exceeded the lower control limit and action limit, indicating that systematic errors occurred and reminding the physicist to take action to improve process performance. The process capability index is an important tool that provides quantitative information used to determine QC limits.
(4) Stacked LSTM models to predict QC records and trends of Linac
This part of the study presents that the stacked LSTM model predict to the QC records and trends of two Linacs. First, the dataset is divided into three sets: the training set was used to train models with different hyperparameter combinations and to combine different sets of hyperparameters using greedy coordinate descent; the validation set was used to determine the best hyperparameters; and the test set was used to evaluate the accuracy under the best hyperparameter combinations. The evaluation criteria included mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R2). Also, the classical time series model ARIMA was applied to compare the performance of stacked LSTM on the same data set. The stacked LSTM and ARIMA models were also used to predict the daily QC data records of another Linac under the same combination of hyperparameters. In the data records, the mean values of MAE, RMSE, and R2 were 0.013, 0.020, and 0.853, respectively, in the stacked LSTM, compared with 0.021, 0.030, and 0.618, respectively, in the ARIMA. The stacked LSTM outperformed the ARIMA for all 23 QC items, with the best prediction was couch rotation (LSTM: MAE = 0.001, RMSE = 0.001 and R2 = 0.975; ARIMA: MAE = 0.002, RMSE = 0.004 and R2 = 0.436); the worst prediction was gantry relative (LSTM: MAE = 0.006, RMSE = 0.007 and R2 = 0.095; ARIMA: MAE = 0.004, RMSE = 0.006 and R2 = 0.383). Overall, the stacked LSTM had better predictive performance than the ARIMA. The trend line lies within the tolerance. The physicist can perform preventive maintenance on the Linac in advance. The stacked LSTM can accurately predict QC records and trends, which is robust.
The methodology used in this study covers only some of the QC items for Linacs, but the methodology can be used as a reference for determining other QC items for Linac, and for determining QC items for other radiotherapy equipment.
[Keywords] linear accelerator; quality control; frequency; limits; failure.