A Comparative Study on the Generalization Ability of back Propagation Neural Network and Support Vector Machine for Tracking Tumor Motion in Radiotherapy

Abstract

Purpose: To compare the generalization ability of back propagation neural network (BPNN) and support vector machine (SVM) in predicting tumor motion. The generalization ability refers to the good prediction on the new data that don't appear in the modeling phase. Methods: The comparison included two aspects: precision and real-time capability. BPNN and SVR were both applied on the same bi-modal liver motion data which were shared on a website. The data consist of target motion and three external skin markers' motion. To simulate the different motion traces in modeling phase and predicting phase, the data of first 2-minute session were used to build the model and used the remaining session for validation. Because the data in the first 2-minute session were collected when the subject breathed freely and the rest contained breathing artefacts. Results: BPNN has a lower root-mean-square error (RMSE) and a lower average prediction time than SVM. Conclusion: In tracking a respiration-induced moving target, BPNN has a better generalization ability than SVM.