MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Lung Tumor Motion Prediction Prior to 4DCT Simulation: Super-Learner Model Development and Clinical Assessment

H Lin*, J Zou , T Li , S Feigenberg , B Teo , L Dong , University of Pennsylvania, Philadelphia, PA

Presentations

(Thursday, 7/18/2019) 10:00 AM - 12:00 PM

Room: 303

Purpose: The aim of this study is to assess a machine learning model for lung tumor motion prediction with a goal for automatic motion management selection. The model performance is evaluated by comparing to an existing clinical protocol.

Methods: Sixteen imaging and eleven clinical features were extracted from non-4D diagnostic CT images and Electronic Health Records (EHR) of 150 consecutive proton patients to characterize lung tumor motion. A super-learner model was built to utilize the input features and combine four optimized machine learning models including the Random Forest, Multi-Layer Perceptron, LightGBM and XGBoost to obtain superior performance. The outputs of the super-learner model consist of motion extent prediction in the Superior/Inferior (SI), Anterior/Posterior (AP) and Left/Right (LR) directions, and were compared against tumor motion extents measured in the free-breathing 4DCT scans. The accuracy of predictions was evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The predicted motion values were then utilized to determine if the patient require active motion management. The current in-house threshold for active motion management is 8 mm. For simplicity, our clinic currently uses a lobe-based method: if the tumor is located in the lower lobe, the patient is assumed to require active motion management such as abdominal compression, breath-hold etc.

Results: The MAE and RMSE of predicted motion extents in the SI, AP and LR direction were 1.2 mm and 1.7 mm, 0.8 mm and 1.2 mm, 0.7 mm and 1.0 mm respectively. Compared to the existing lobe-based method, the super-learner model improved the prediction of large tumor motion (>=8mm) by 9%, and improved the prediction of small tumor motion (<8mm) by 30%.

Conclusion: Our findings indicate the feasibility of developing a super-learner model to derive accurate tumor motion prediction, and could simplify the motion management workflow prior to the 4DCT simulation.

Funding Support, Disclosures, and Conflict of Interest: Dr. Lei Dong is in the Speaker Bureau for Varian Medical Systems. Dr. Taoran Li has a speaker agreement with Varian Medical Systems unrelated to this work.

Keywords

Not Applicable / None Entered.

Taxonomy

TH- External Beam- Particle therapy: Proton therapy - motion management(intrafraction)

Contact Email