Click here to

Session: Artificial Intelligence in Treatment Planning and Delivery [Return to Session]

Predicting Daily Record of Machine Performance Check Using Multi-Long Short-Term Memory Neural Networks

M Ma*, J Dai, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, China, BeijingCN,


SU-A-TRACK 6-4 (Sunday, 7/25/2021) 10:30 AM - 11:30 AM [Eastern Time (GMT-4)]

Purpose: The test of machine performance check (MPC) monitors the status of Varian truebeam machine. The cumulative record of MPC tests shows the trend of machine performance and provides warning message for preventive actions. To precisely predict the future records of MPC tests, a generalized model using multi-long short-term memory (multi-LSTM) was introduced.

Methods: A total of 1044 sets of daily MPC test data were collected. The 80% of data were used to train multi-LSTM model, while the remaining 20% of data were used to test the accuracy of the model. To enrich the data, cubic interpolation is used in the model. Furthermore, a greedy coordinate descent method was employed to find the optimal hyper-parameters of the model. The accuracy of the model was quantified by the mean absolute error (MAE), root‐mean-square error (RMSE), and coefficient of determination (R2).

Results: The hyper-parameters are determined as 15 for the length of time lags, Adam for the optimizers, 0.001 for the learning rates, 150 for the number of epochs, 50 for the number of hidden units, and 32 for the batch sizes. The majority of MPC test records was predicted with an RMSE less than or equal to 0.005. Exceptions are the jaw collimation, collimation rotation offset, beam output change, and beam uniformity. The similar result was observed in MAE metric. The result of all MPC tests records with R2 was higher than 85.00%.

Conclusion: The multi-LSTM model can accurately predict all MPC test record. Predicting the future performance records of MPC tests will foresee possible machine failure, which allows early action of machine maintenance and reduces unscheduled machine downtime.

Funding Support, Disclosures, and Conflict of Interest: the National Natural Science Foundation of China (Grant No. 11875320)



    Quality Assurance, Quality Control, Modeling


    IM/TH- Mathematical/Statistical Foundational Skills: Machine Learning

    Contact Email