Adaptive Confidence Regions Of Motion Predictions From Population Exemplar Models

Precise radiation therapies require not only accurate prediction of the motion of the structures in the treatment region, but also confidence values of these predictions to enable planning of residual motion and detection of failure predictions. While various motion models have been proposed for the prediction of motion in the abdomen due to free-breathing, none has provided confidence regions. In this study we use the conditional probability density function of statistical liver motion models for predicting confidence regions, propose a method for optimizing the accuracy of the confidence regions and show the adaptability of the confidence regions due to partial observations when using exemplar models. The average accuracy of the confidence regions of single Gaussian (SG) models could be improved to the level of the exemplar models. Exemplar models provided on average better motion predictions (1.14 mm) and slightly smaller 68% confidence regions (1.36 mm) than the SG models (1.21 mm, 1.43 mm resp.). The confidence region size correlated temporally on average weakly (r = 0.35) with the errors of the motion prediction for the exemplar models, leading to a higher percentage of treatable locations and lower motion prediction errors per duty cycle than SG models.