Program Information
Radiomics-Based Treatment Outcome Prediction Based On Belief Function Theory and Sparsity Learning
Z Zhu1 , T Mazur1 , C Lian2 , S Ruan2 , M Anastasio1 , J Wu1 , B McClain1 , J Williamson1 , P Grigsby1 , S Mutic1 , H Li1*, (1) Washington University School of Medicine, Saint Louis, MO, (2) University of Rouen, Rouen, Normandy, France
Presentations
TH-AB-201-12 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201
Purpose: The suboptimal predictive performance of currently available radiomics-based treatment outcome prediction methods is due, in part, to inadequate management of the redundancy, heterogeneity, and uncertainty of features. We developed a novel model based on belief function theory (BFT) and sparsity learning to address these issues of radiomic features for accurate and reliable outcome prediction.
Methods: The proposed model includes feature preparation, predictive feature subset selection, and outcome prediction. First, complementary and quantitative radiomic features, including shape-based, intensity-based, histogram-based, and perceptual and wavelet-transform based textural features are first extracted from each patient case. Then, a BFT-based method is designed to find a sparse predictive feature subset from all extracted features through the minimization of a loss function using training samples with known outcome classes. The loss function is defined by combining errors of the classification result, penalties for un-predictive features and sparsity regularization to find the best low-dimensional transformation of the original feature set. Finally, an evidential K-nearest neighbor (EK-NN) classifier is used to predict the outcome of a patient case given the selected predictive feature subset as the input.
Results: We have compared our model with other reported methods on 63 stage II-III cervical cancer patients, 25 lung cancer patients, 36 esophagus patient cases, and 45 lymphoma patients. The prediction accuracy and AUC of our model are 0.94/0.98 on cervical cases, 0.94/0.94 on lung cases, 0.83/0.82 on esophagus cases, and 0.93/0.92 on lymphoma cases, which demonstrates feasibility and the superior performance of our model compared to other reported methods.
Conclusion: Our model addresses the difficulties of using radiomic features in a unified framework and seamlessly integrates feature extraction with BFT-based feature selection with a new EK-NN classifier for prediction. This model has the potential to guide patient-specific treatment and advance the state of science in personalized radiation therapy.
Contact Email: