Click here to

Session: Multi-Disciplinary: Biologically and Functionally-Guided Radiation Therapy [Return to Session]

Machine Learning Based Prediction of Big Inter-Fractional Changes in Lung Tumor Motion During Radiotherapy Course Using Robust Feature Selection

Q Wang*, L Dong, B Teo, H Lin, S O'Reilly, University of Pennsylvania, Philadelphia, PA


WE-IePD-TRACK 3-3 (Wednesday, 7/28/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: During lung cancer radiotherapy, tumor motion amplitude may change due to various reasons. If motion amplitude varies greater than a clinical threshold (3 mm was used in this study), adaptive radiotherapy may be needed. In this work, we proposed an ensemble machine learning model to identify potential patients for adaptive radiotherapy after 4DCT simulation.

Methods: We employed 5 conventional classification models (logistic regression, decision tree, support vector machine, random forest, and XGBoost) as base predictors. Then the results were used as inputs for a two-layer neural network to get a final prediction. For each first-level predictor, the inputs are based on 17 features (clinical factors and imaging features from planning CT) from 140 patients. To understand the correlation between each feature and the prediction task, we proposed a robust feature ranking method. We compared the ensemble learning model with each individual base predictor. To evaluate the model robustness, the error and area under receiver operating characteristic (ROC) curve were derived based on 1000-times 5-fold cross validation.

Results: The feature ranking results showed that the initial tumor motion along the superior-inferior (SI) direction best correlates with the big tumor motion change. By sequentially removing the least important features, the prediction performance of base predictors was consistently improved. Furthermore, the proposed ensemble learning model effectively improved the best predictive accuracies of a single base predictor (Random Forest or Logistic Regression) from 0.78 to 0.83, and the best F1 score of XGBoost from 0.50 to 0.60.

Conclusion: The proposed ensemble learning model can effectively detect a large change in tumor motion, which helps predict patients that may need adaptive radiotherapy. Experimental results demonstrated superior predictive performance over 5 base predictors. By involving a feature ranking strategy, the prediction accuracy of both base predictors and the overall prediction model are improved.

Funding Support, Disclosures, and Conflict of Interest: This research was in part sponsored by a research grant from Varian Medical Systems.



    Statistical Analysis, Lung, Feature Selection


    TH- External Beam- Photons: Motion management - interfraction

    Contact Email