Purpose: An unsupervised machine learning method, autoencoder (AE), was applied to error detection of treatment plans in radiotherapy.
Methods: Treatment plans of more than 500 breast cancer patients between 2010 and 2020 were collected from MOSAIQ system and analyzed retrospectively. Those plans were hybrid intensity-modulated radiotherapy (IMRT) plans with two tangential conformal fields and two tangential intensity-modulated fields. The most important 9 parameters were extracted from plan for each patient. A classical autoencoder (AE) method was used to encode the original features to fewer features and then reconstruct the original features from the encoded features. The average mean-square error (MSE) between the reconstructed features and the original features for all training data was calculated and set as baseline for error detection. For a new patient, the 9 features were extracted from treatment plan and fed into trained AE model. The reconstruction MSE was then calculated by AE and compared to the baseline value for the difference. If the difference is more than a threshold value, this treatment plan is considered as an outlier and need further checking. In addition, errors in various levels were purposely introduced to each parameter in testing dataset. The resulting error detection rates were used to evaluate the sensitivity of AE model.
Results: AE model detected all errors with 100% detection rate in testing dataset. For sensitivity analysis, 1 parameter was not affected as its value always fixed to two constants. 4 parameters were less sensitive to errors and another 4 parameters were more sensitive to errors. For all tested parameters, the false positive rate is about 0.01%.
Conclusion: AE method provides a novel way for automatic treatment plans checking, which is more efficient and robust.