Detection of suboptimal IMRT treatment plan using machine learning on radiomics features of dose distribution for lung cancers

Quality assurance in radiotherapy is an important process so that the use of radiation provides maximum ben-efits. Currently, the implementation of machine learning (ML) in the quality assurance of treatment planning is growing. In this study, 34 optimal intensity-modulated radiation therapy (IMRT) treatment plans and 10 sub-optimal IMRT treatment plans obtained from Siloam MRCCC Semanggi Hospital were used to train an ML model of the type autoencoder for anomaly detection developed using PyTorch. There were four stages in this study, namely the preparation stage, development stage, validation stage, and evaluation stage. At the development stage, the raw data was prepared so that it is ready to be used for training. At the development stage, the model was developed and a hyperparameters optimization was performed. The accuracy of the model was analyzed at the validation stage. Finally, at the evaluation stage, the model performance was evaluated along with a Mann-Whitney U test on dose-volume histogram (DVH) parameters, radiomics features, and DVH metrics (conformity index and homogeneity index) to show the difference between treatment plans. The model used 161 radiomics features with an epoch of 1250 iterations, 150-50-17 hidden layers configuration, and a learning rate of 0.2 as the most optimal configuration. The results showed an accuracy of 36% with 7% of radiomics features, 50% of DVH parameters, and the homogeneity index being different significantly. After refinement, that is removing data with conformity index below one, the accuracy became 50% with 12% of radiomics features, 45% of DVH parameters, and both DVH metrics being different significantly. If the radiomics features used are those that were significantly different, the accuracy increased to 93%. From these results, it can be concluded that most radio-mics features are noise that need to be removed so the model can detect suboptimal treatment plans. In addition, the planning target volume segment along with the firstorder radiomics features group is the main differentiator between the optimal and suboptimal treatment plans.

Automated summary

This study used machine learning to train a model for anomaly detection in the quality assurance of radiotherapy treatment planning. The results showed that most radiomics features are noise that need to be removed for the model to detect suboptimal treatment plans. The planning target volume segment along with the first-order radiomics features group is the main differentiator between optimal and suboptimal treatment plans.