Room: Exhibit Hall
Purpose: Treatment uncertainty due to respiratory motion may alter the underlying dose distribution for a patient receiving proton liver treatment. The process of obtaining a 4D treatment plan is however complicated and time consuming. The purpose of this study is to apply the support vector machine (SVM) algorithm to predict whether difference between a 3D and a 4D dose plan is small enough on a field-by-field basis to allow omission of the resource-intensive 4D planning procedure.
Methods: Totally 77 single-field treatment plans from 37 patients treated in our institution were included in this planning study. For each field, 3D treatment planning (based on average CT) and 4D treatment planning were performed respectively. The 4D plan was obtained by applying the same 3D beam setup, calculating dose on each CT image of the 4DCT breathing phases and summing up doses after dose deformation based on deformable vector fields from deformable registration.The input features extracted from CT images and 3D treatment planning were imported into the MATLAB â€œClassification Learnerâ€? application to train the SVM classifier for critical organs including heart, normal liver and bowel.Receiver operating characteristic (ROC) curve was used to evaluate the performance of the SVM classifiers.
Results: The SVM prediction classifiers for heart mean dose, normal liver mean dose and bowel V34Gy showed sensitivity of 100%, 70% and 87.5%; specificity of 95%, 70.59% and 68.42%; area under the ROC curve (AUC) of 0.9643, 0.6824, and 0.7895 respectively.
Conclusion: We have successfully developed a model using the SVM classifier to predict whether dose difference for selected critical organs between a 3D and a 4D dose plan is significant enough to warrant a 4D planning evaluation.