Program Information
Feasibility Study of a Feature Based Prediction for Optimal Position Selection for Left-Sided Breast Radiotherapy
X Tang1*, H Lin2 , T Liu2 , C Shi3 , S Petillion4 , I Kindts4 , X Xu2 , (1) Memorial Sloan Kettering Cancer Center Westchester, West Harrison, NY. (2) Rensselaer Polytechnic Institute, Troy, NY. (3) Memorial Sloan Kettering Cancer Center, Basking Ridge, NJ. (4) Universitair Ziekenhuizen Leuven, Leuven, Gasthuisberg, Belgium.
Presentations
MO-RAM-GePD-J(A)-6 (Monday, July 31, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: A left-sided breast cancer patient might receive lower OAR dose in supine free breathing (FB), supine Deep Inspiration Breath Hold (DIBH), or prone free breathing (prone). Physicians often make the decision based on experiences and trials, which might not always be optimal. We herein conduct a feasibility study of a feature based prediction mathematical model to predict the lowest OAR doses among these three positions, providing a quantitative tool for corresponding clinical decision.
Methods: Patients were scanned in FB, DIBH, and prone positions under an IRB approved protocol. Tangential beam plans were generated for each position, and OAR doses were calculated. The position with least OAR doses is defined as the optimal position. Ten anatomical features were extracted from each FB scan to build the model. We performed univariate feature selection and recursive feature elimination to outweigh the most important features. The top five features were selected using the ranking method in recursive feature elimination to determine feature importance, and these features are then fed into nine machine learning models as inputs and used to train the models. The statistical modeling was performed using SciKit-Learn.
Results: The top five important features selected by the recursive elimination includes: breast volume, heart volume, laterality of heart, Distance between ipsilung and breast, in-field heart volume. Leave-one-out cross-validation and area under Receiver Operating Characteristic (AUROC) were used for the model evaluation. The best classification performance was Support Vector Machine with radial basis function with 0.8571 accuracy, and the AUROC was 0.6999±0.21.
Conclusion: The proposed feature selection-based predictive model is feasible for the optimal positioning selection of left-sided breast irradiation.
Contact Email: